OECD Digital Economy and Society Outlook 2026

Report

5 November 2026

Spotlight. The innovation landscape of privacy-enhancing technologies

Copy link to Spotlight. The innovation landscape of privacy-enhancing technologies

Abstract

This Spotlight examines the fast-evolving innovation landscape of privacy-enhancing technologies (PETs), which are tools designed to protect confidentiality and privacy while enabling responsible data use and sharing. The latest generations of PETs hold growing economic and strategic significance as they increasingly contribute to reconciling national, organisational and individual control over data to support data sharing and cross-border collaboration. Despite growing recognition of their potential in critical domains and sectors, adoption of PETs remains limited due to perceived complexity, low awareness, unclear return on investment and regulatory uncertainty, particularly for small and medium-sized enterprises. Drawing on new quantitative indicators, the Spotlight presents comparative evidence on PET-related scientific publications, open-source software developments and patenting activities. It complements these supply-side trends with analysis of demand-side considerations, including emerging use cases and policy initiatives. In so doing, it offers a first empirical glance at the global PET ecosystem and its evolving drivers of innovation and adoption.

Introduction

Copy link to Introduction

Privacy-enhancing technologies (PETs) are digital technologies that enable the collection, processing, analysis and sharing of information while protecting personal or otherwise confidential data (OECD, 2023_[1]). Many technologies considered PETs were first developed decades ago in academic or theoretical settings. However, recent advances in cryptography, computing power and network connectivity have made them deployable in real-world systems at scale.

These advances open new opportunities for data collaboration that were previously infeasible or carried disproportionate confidentiality, privacy or security risks. For instance, PETs now make it possible for multiple parties to jointly train or share artificial intelligence (AI) models without transferring the underlying data (OECD, 2025_[2]). Accessing new data sources while ensuring compliance with privacy and data protection and other obligations is increasingly relevant: organisations face rising data quality challenges during the development of their AI systems (OECD, 2025_[3]; 2025_[4]).

Beyond privacy, PETs can also protect intellectual property and commercially or otherwise sensitive information. This protection reduces disclosure and digital security risks that can limit collaboration and undermine trust in the face of escalating reputational and financial costs of digital security incidents. One major study estimates that the average global cost of a data breach was USD 4.44 million in 2025, up from USD 3.79 million in 2015. The healthcare sector alone averaged USD 7.42 million per breach with the compromise of companies’ intellectual property being the costliest type of record breached (IBM, 2025_[5]).

PETs thus hold significant economic and strategic value in addition to their privacy and confidentiality benefits. They help balance data control with the benefits of sharing and cross-border co-operation, especially as concerns about security and data sovereignty increase. By enabling trust in the global data ecosystem, PETs can thus support AI-driven innovation and global value chains. In this sense, they can be seen as key technologies enabling data free flow with trust (DFFT).

Despite their growing technical maturity and demonstrated potential for addressing economic, policy and social challenges, however, the deployment of PETs across sectors and organisations remains limited. Many organisations still perceive PETs as too complex or costly to integrate into existing data infrastructures. This perception became apparent from OECD expert workshops,1 stakeholder consultations in diverse sectors and jurisdictions, and insights provided by the OECD DFFT Experts Community Working Group on PETs.

These barriers, in turn, have detrimental effects on the most innovative PET service providers. Many of these providers are start-ups or small and medium-sized enterprises (SMEs) that face high upfront customer acquisition costs. Each new client typically requires substantial engagement and demonstration to understand the value and applicability of PETs. This Spotlight seeks to address these problems, identifying the structural constraints that impede uptake of PETs, while highlighting how countries are shaping data governance policies to advance their adoption.

This Spotlight describes the economic landscape of PETs, focusing on innovation and policy trends that underpin development and shape adoption from both a supply- and a demand-side perspective. On the supply side, it examines patterns across the innovation lifecycle – from scientific research and open-source software (OSS) development to patenting. On the demand side, it considers emerging PET use cases and the policy, regulatory and market factors that influence adoption across sectors and countries. These include data protection, privacy regulations and more targeted policy measures, which appear to have contributed to PET-related innovation and adoption. Broader opportunities to analyse personal data in the context of data-driven innovation can also drive growth in PETs. However, analysis of these incentives falls beyond the scope of this paper.

By mapping these dynamics, this Spotlight presents first-of-a-kind empirical insights into the economic and innovation landscape of PETs. It shows that the PET ecosystem demonstrates clear signs of technological maturity and growing strategic importance. However, it also illustrates how the broad diffusion (e.g. in research applications) and commercial deployment of PETs remain constrained by a range of structural and regulatory challenges.

Addressing these barriers is of strategic importance to policymakers, particularly in the context of international co-operation. As such, several areas are identified for deeper analysis of the innovation, market and policy factors shaping their development and adoption. This forthcoming analytical work will support the set-up of a Global Repository of PET Use Cases at the OECD. The repository will provide an evidence base to support greater alignment of regulatory and policy practices around PETs, including in cross-border use cases that directly contribute to DFFT.

Research, open-source development and patent filings in PETs have surged since 2015, partly driven by regulatory incentives

Copy link to Research, open-source development and patent filings in PETs have surged since 2015, partly driven by regulatory incentives

Like other advanced digital technologies, the development of PETs takes place within a complex innovation ecosystem spanning the full innovation cycle – from fundamental research to experimental implementation to commercial application (OECD, 2015_[6]). Analysing trends across these stages helps trace how ideas move from early theoretical work to product development (including software implementation) and eventual market protection through intellectual property (e.g. patents). This, in turn, offers insights into global market and innovation dynamics that shape technology adoption.

Combined analysis of global data on PET-related scientific publications, OSS projects and patent applications between 2010 and 2025 reveals a striking acceleration across these phases of the PET innovation cycle.2 Although innovation rarely follows a strictly linear path (OECD, 2020_[7]), the evidence indicates a coherent progression: research output expands alongside subsequent implementation efforts, eventually followed by formal protection through patents. Although increasingly prominent, OSS development and patenting are only two potential routes from concept to market deployment. The choice of route can depend on firms’ market strategies and countries’ patent systems. Other forms of knowledge diffusion such as copyright, trade secrets or open research collaborations, including between public research and innovative entrepreneurs (Breschi et al., 2019_[8]), remain outside the scope of this quantitative analysis. However, they are also central to the PET innovation ecosystem.

Since 2015, research, OSS development and patenting activity in PETs have expanded in parallel, although at different speeds (Figure 2.S.1). Academic publications on PETs grew 20-fold by 2025 (reaching an index of 2 005), while open-source repositories increased about 13-fold (index 1 307) (see Annex 2.S.A and Annex 2.S.B for information on the methodology). Patent filings also rose steadily, almost fourfold between 2015 and 2022 (index 355), reflecting the release of mature research ideas for their commercialisation and legal protection. The annual time series suggest parallel acceleration, where peaks in implementation often coinciding with rapid research expansion. Meanwhile, patent filings typically appear with a two-to-three-year delay as technologies approach market maturity.

Figure 2.S.1. Trends in patent applications, scientific publications, and OSS development show that innovation in PETs is surging
Copy link to Figure 2.S.1. Trends in patent applications, scientific publications, and OSS development show that innovation in PETs is surging

*Growth of PETs in patent applications, scientific publications, and OSS development (Index 2015=100), 2010-2025*

The cumulative and reinforcing nature of the PET innovation cycle is illustrated by the trajectory of federated learning, originally prompted by Google’s reference publication in 2016 (McMahan et al., 2016_[9]). This research catalysed a surge of academic work on federated learning (around 1 750 papers by 2021), which in turn led to the release of open-source frameworks such as Google’s TensorFlow Federated (2019). Another example is the decades-long evolution of homomorphic encryption – from theoretical concept (Gentry, 2009_[10]) to practical software repositories (Microsoft, 2026_[11]) to commercial deployment.

In both cases, large digital firms acted as ecosystem anchors, with their respective software repositories enabling early commercial deployment and spurring further innovation. This pattern is reflected in the OSS ecosystem, where large technology companies dominate both by the number of repositories and visibility metrics, measured by the total GitHub stars attributed to each repository (Figure 2.S.2).3 However, the sample size is relatively small, and GitHub use reflects only a partial picture of PET software development. Together, this means these figures should be interpreted cautiously and serve as conservative estimates of commercial activity in the OSS ecosystem.

Figure 2.S.2. A large range of corporate entities are active in the OSS ecosystem
Copy link to Figure 2.S.2. A large range of corporate entities are active in the OSS ecosystem

*Top businesses by GitHub PETs repository count (Panel A) and by total stars (Panel B), February 2026*

The surge in the number of firms active in developing PETs is illustrated by the growth in PET-related patent applications from private companies (Figure 2.S.3). In 2010, 707 unique companies filed a PET-related patent for the first time. By 2022, this number had increased by almost fourfold to 2 585, underlining the rise in the strategic development of PETs as intellectual property leveraged in commercial applications. However, not all innovations follow the same research-to-patent pathway. Many companies, such as in blockchain (distributed ledger technologies), prioritise OSS release over patent protection. As such, they view community adoption and network effects as more valuable than holding intellectual property rights. In contrast, some organisations pursue proprietary development without open-source release, maintaining competitive advantage through trade secrets rather than patents. These alternative pathways explain apparent disconnects between research volume and patent/repository counts in certain types of PETs and certain sectors (see next sections).

Figure 2.S.3. The number of new firms patenting PETs is on the rise
Copy link to Figure 2.S.3. The number of new firms patenting PETs is on the rise

*Number of unique firms filing PET-related patents for the first time, 2010-2022*

The simultaneous surges across all phases of the PET innovation cycle between 2015 and 2025, marked by a 20-fold increase in research output, 13-fold growth in OSS development and a 4-fold rise in patent filings, coincides with several related trends. These trends include the diffusion of data-driven digital technologies and data-intensive business models (e.g. AI, automated decision systems, ubiquitous connectivity, cross-border cloud infrastructure), and escalating cybersecurity threats (e.g. data breaches). It also coincides with the introduction and entry into force of laws and regulations concerning data protection and privacy, cybersecurity, health, online child protection, etc., all around the globe. Taken together, these frameworks converge towards stronger safeguards for data processing, interoperability and sharing, and raise liability and reputational risk for data breaches or misuse.

In the European Union (EU), a turning point was the adoption of the General Data Protection Regulation (GDPR). Adopted in 2016 and effective from 2018 (European Union, 2016_[12]), the GDPR led research funding bodies to explicitly link PET research projects to GDPR compliance and Article 25 on “data protection by design and by default”. This trend was reinforced by subsequent legislation such as the California Consumer Privacy Act, introduced in 2018 and entering into force in 2020 (State of California, 2018_[13]), which has since amplified similar compliance incentives across other major digital markets. Examples from the People’s Republic of China (hereafter “China”) include the Cybersecurity Law (Cyberspace Administration of China, 2016_[14]), the Personal Information Protection Law (National People’s Congress of the People’s Republic of China, 2021_[15]) and the Data Security Law (National People’s Congress of the People’s Republic of China, 2021_[16]). These entered into force in 2017, 2021 and 2021, respectively.

Generative AI and blockchain drive the current wave of research and innovation on PETs

Copy link to Generative AI and blockchain drive the current wave of research and innovation on PETs

The intersection of AI and PETs represents perhaps the most transformative trend in the field, shaping and accelerating research and development (R&D) on PETs. As shown in Figure 2.S.4, a significant share of scientific papers and OSS repositories are related to PETs and AI. While federated learning is the most obviously AI-related PET, it is not the largest source of recent activity. Synthetic data dominate by research volume and are growing fastest, featuring in the region of 3 000 papers in 2025. They have been experiencing steady growth in OSS implementation through a mix of AI and non-AI tools (e.g. generative model toolkits vs. test-data generators and masking utilities). Differential privacy also shows a substantial uptake.

Figure 2.S.4. Research and development on PETs is partly driven by AI
Copy link to Figure 2.S.4. Research and development on PETs is partly driven by AI

*Percentage of PETs related to AI applications across federated learning, synthetic data and differential privacy, 2015-2025*

Evidence reveals a clear divergence between PET research and implementation.4 Publication counts continue to rise across all major types of PETs, while OSS activities have stagnated or even declined for federated learning, possibly indicating a phase of consolidation. Research on synthetic data and differential privacy is substantially more AI-oriented than OSS implementation. Papers on synthetic data are predominantly focussed on generative AI, whereas software repositories still have a large non-AI focus such as on masking tools. The inversion is most striking for federated learning. Here, research outputs remain above the 2024 peak, yet OSS implementations have fallen well below 2024 levels. However, newly created repositories can take time to be discovered and forked, which means that discovery lag may partially explain the observed gap.

Yet these AI-related PETs are likely to remain highly attractive as demand for privacy-preserving data sharing and model training continues to grow. AI systems rely increasingly on high-quality and often sensitive data combined with emerging AI and data governance obligations. This creates strong incentives for these PETs (OECD, 2025_[3]; 2025_[4]), as does the increase in the number of reported AI incidents and hazards (OECD-GPAI, 2026_[17]; OECD, 2026_[18]).5 More importantly, as PET capabilities mature, they increasingly make new categories of use cases both technically and legally feasible, including cross-border data collaborations previously considered too risky or infeasible (see Box 2.S.1).

Box 2.S.1. Using privacy-enhancing technologies to enable international data sharing

Copy link to Box 2.S.1. Using privacy-enhancing technologies to enable international data sharing

The United Kingdom (UK) Department for Science, Innovation and Technology (DSIT) piloted the use of PETs to enable the National Disease Registration Service of the National Health Service (NHS), alongside the US National Cancer Institute, to securely study ultra-rare childhood tumours, while keeping NHS data under NHS control. The pilot, building on the 2022-2023 UK-US PETs prize challenges, demonstrates how PETs can advance digital innovation and deliver on the UK Blueprint for Modern Digital Government.

The data-scarcity problem: each childhood cancer subtype is extremely rare – some affect fewer than two children per million annually. Pooling data internationally helps clinicians and researchers, but privacy and regulatory constraints make exporting patient-level data slow, costly or impossible. DSIT identified the need for “a better technical approach”.

A PET-led solution: using the Duality platform, the pilot “moved the code, not the records”. Researchers could perform joint analyses without transferring personal data outside NHS systems. The solution combined:

Federated querying within each partner’s firewall, returning only site-level aggregates;
Trusted execution environments (TEEs) to combine results in a secure enclave;
Differential privacy and cell suppression adding noise and hiding results; and
Embedding privacy preservation into the architecture to reduce information governance friction.

Outcomes and lessons: analysts ran no-code queries from a central interface. The PET workflow cut the governance process from over a year to about two months, allowing more time for research while maintaining rigorous security checks. DSIT noted that guidance from the UK Information Commissioner’s Office on PETs was instrumental.

Key lessons for government:

PETs turn information governance into a design choice, not an afterthought.
Trust modelling beats trust transfer – TEEs let partners verify code execution locally.
Parameter transparency matters for governance evaluation.
Iterate with dummy or synthetic data first to align analysts and governance teams.

Implications: this UK-US pilot shows that PETs can make international data collaboration faster, safer and more trustworthy, turning privacy compliance into an enabler of responsible, cross-border research.

Source: Durkee et al. (2025_[19]), Using privacy enhancing technologies to enable international data sharing, https://gds.blog.gov.uk/2025/10/09/using-privacy-enhancing-technologies-to-enable-international-data-sharing.

Alongside AI-driven developments, blockchain-related PETs have emerged as another major engine of applied innovation, particularly through the development of zero-knowledge proofs and related smart-contract frameworks.

Combined analysis of scientific publications and GitHub repositories shows that blockchain is the only domain in which implementation activity far exceeds research output in PETs.6 There are more than 550 active OSS repositories compared to around 150 academic papers, a near 4-to-1 ratio. This inversion reflects the commercial incentives of decentralised finance and digital-asset markets, where open-source and rapid iteration are perceived as key success factors within the blockchain community (Canidio, 2018_[20]). Consequently, blockchain has become a leading testbed for real-world deployment of PETs in finance.

Health and finance are the most active application domains for PETs

Copy link to Health and finance are the most active application domains for PETs

Analysis of PET activity across application domains shows clear sectoral concentrations, with healthcare and finance standing out as the most active (Figure 2.S.5). Yet their innovation profiles differ markedly: healthcare features in 13% of all PET-related papers compared to 6% of PET-related GitHub repositories. This suggests that PET activity is concentrated relatively similarly across research and OSS implementation. Finance, however, is significantly more prominent in OSS implementation, comprising 31% of PET-related GitHub repositories and only 7% of research papers.

Figure 2.S.5. The uptake of PETs varies across application domains
Copy link to Figure 2.S.5. The uptake of PETs varies across application domains

*Percentage of publications and repositories of each PET attributed to each application domain, 2010-2025*

Healthcare and medical research publications have a strong emphasis on synthetic data and differential privacy, technologies that enable secondary data use while preserving patient privacy (Durkee et al., 2025_[19]; OECD, 2025_[21]). Research efforts concentrate on clinical applications and multi-site collaborations in areas like brain imaging, cancer prediction, patient monitoring and epidemiological surveillance, spurred by the COVID-19 pandemic. In terms of OSS repositories, a few mature initiatives such as Kaapana (a medical-imaging platform) account for the small share of roughly 180 healthcare-related OSS repositories, compared with more than 930 in finance.

Finance and banking represent the highest relative share of PET-related research activity and OSS development, accounting for about 70% of all zero-knowledge proof activity and 60% of secure multi-party computation (MPC). This reflects the central importance of cryptography-enabled privacy in blockchain-based applications (ConsenSys, 2026_[22]; Ben-Sasson et al., 2014_[23]), where blockchain-focussed companies are highly active in OSS repositories. At the same time, more traditional financial use cases such as those supporting efforts to combat money laundering and promote regulatory compliance remain prominent in scientific literature (BIS Innovation Hub, 2023_[24]; Mastercard, 2024_[25]).

A broader comparison across application domains reveals distinctive application patterns. Database and data-management systems show the greatest focus on anonymisation techniques (around 31% of related PET activity), while cloud-computing environments concentrate on trusted execution environments (TEEs) and homomorphic-encryption solutions (roughly 43% and 22%, respectively), reflecting the need to preserve confidentiality in shared infrastructures. By contrast, the Internet of Things (IoT) displays a more fragmented pattern, with no single PET type exceeding 21% of total activity.

Within this landscape, industrial and manufacturing applications represent a smaller but distinct pattern of PET adoption, driven primarily by non-personal data protection needs. Activity is growing in areas such as supply-chain confidentiality, predictive maintenance, and federated learning for quality control. These are areas where confidentiality concerns centre on proprietary process data, sensor telemetry and trade secrets. Adoption remains exploratory, spanning all types of PETs. OSS implementation, however, remains extremely limited (at 1% of repositories). This may reflect a preference among major industrial players for proprietary solutions, often protected through patents rather than shared as OSS tools.

International patent filing trends reveal divergent national concentrations and specialisations in PETs between 2010 and 2022

Copy link to International patent filing trends reveal divergent national concentrations and specialisations in PETs between 2010 and 2022

Over the past decade, patenting activities in PETs have expanded rapidly – from 1 455 applications filed in 2010 to 10 729 in 2022. Since national patent systems differ in coverage and incentives, data on patenting activity are not directly comparable across countries. However, they do capture the overall scale and dynamics of inventive activities. To assess cross-border trends and international innovation capacities, the analysis that follows focuses on international patent families, i.e. inventions filed in more than one jurisdiction or through the Patent Cooperation Treaty. It applies fractional inventor counts in line with common OECD methodologies (OECD, 2009_[26]; 2025_[27]).7

Between 2010 and 2022, the number of international PET patent families increased from 396 to 2 344. When nationally filed patents are included, the overall volume is substantially larger, as only around 40% of PET patents are filed internationally. This compares to around 35% for information and communication technology (ICT) patents overall. This may suggest that some PET inventions are still at a research or early commercialisation stage. They are often developed by academic or specialised actors or by large multinational firms without immediate intentions for global commercialisation.

Within the wider innovation landscape for ICTs, PETs remain a relatively small but fast-growing component. In 2022, the last complete year of data, PETs accounted for 9% of international ICT patent families compared with just 3% in 2010 – a threefold increase in relative weight over little more than a decade.8 This sustained expansion underscores the accelerating importance of PETs as a distinctive field of digital innovation.

The total number of international PET patent families increased from 396 in 2010 to 2 344 in 2022, a near sixfold increase over the period (Figure 2.S.6). This trend was experienced both by OECD Member countries and partner economies, including China. The fall in the total between 2021 and 2022 mirrors the broader slowdown observed across global innovation systems during the COVID-19 pandemic. This can be explained by backlogs in patent offices and a short-term shift in corporate R&D budgets (Fink et al., 2022_[30]; WIPO, 2022_[28]; 2023_[29]; Trunschke et al., 2024_[31]). Performance varied significantly across OECD Members, however. Fourteen countries, including Germany and Canada (which are relatively large contributors), grew their international filings on PETs; five countries experienced moderate declines of below 10%, including the United States, Japan and the United Kingdom; while eight countries faced steeper declines, notably France, Sweden and Switzerland.

Post-2020, international PET patent filings experienced a broad slowdown linked to COVID-19 pandemic effects, patent office backlogs and shifting R&D budgets. Despite these shocks, growth in China’s international filings remained resilient between 2020 and 2022 at +32.5%, with a compound annual growth rate (CAGR) of +15.1%. This was nearly three times the OECD average (+11.1%, CAGR +5.4%), and almost double that of the United States (+17.4%, CAGR +8.4%).9

In contrast to the decline in international filings, domestic (national-only) PET patent filings grew substantially, increasing by 30% globally between 2020 and 2022. This growth was highly concentrated outside of OECD Member countries, primarily in China, which saw domestic filings surge by almost 40% (from around 4 500 to around 6 500 fractional counts). Domestic filings of OECD Members grew more slowly (+7%), with significant heterogeneity between countries. Korea saw strong domestic growth (41%); the United States and United Kingdom remained stable (3-5%); and Japan, Germany and Australia experienced notable declines (between -9% and -24%). That said, national patent filings, including numbers on China, should be interpreted with caution.10 International patent families signal inventions deemed valuable enough for multi-jurisdictional protection and reflect genuine cross-border innovation capacity. However, domestic filings may include lower-quality patents and strategic filings for subsidy purposes. Still, they also show inventions targeting local markets, and when combined with international filings, they capture the overall scale and dynamics of inventive activities.

Figure 2.S.6. PETs patenting is highly concentrated in a handful of economies
Copy link to Figure 2.S.6. PETs patenting is highly concentrated in a handful of economies

*Fractional patent count for top five countries, OECD and World, 2010-2022*

In 2022, the distribution of international PET filings reveals high concentration in a few economies. China, Japan and the United States account for almost two-thirds of all inventor-attributed international patent families. The United States alone represents almost one-third of the total (715 fractional counts of international PET families), followed by China (475) and Japan (216). Korea, Germany and Canada follow, while emerging contributors include India, Israel and Singapore. These data reflect where inventive activity takes place, not where patents are owned or exploited commercially. National innovation systems exhibit distinct specialisation patterns across PET types, as revealed by analysis of the top 15 countries in terms of international PET patent families (Figure 2.S.7).

Figure 2.S.7. PETs specialisation differs by country
Copy link to Figure 2.S.7. PETs specialisation differs by country

*Percentage of PET-related patent fractional counts by top 15 countries, 2010-2022*

Three economic profiles for PET development emerge from this analysis:

Diversified portfolios: countries such as Singapore, Switzerland, India and Japan maintain balanced portfolios spanning several PETs. Singapore shows the broadest PET distribution, with no single one exceeding 25% of its portfolio. Switzerland combines relatively strong positions in TEEs (23%), homomorphic encryption (23%) and secret sharing (25%), while Japan exhibits an even spread across TEEs (25%), anonymisation (23%) and secret sharing (27%). Meanwhile, India’s focus lies in a combination of TEEs (26%), federated learning (23%) and anonymisation (21%).
Moderately diversified portfolios: these economies demonstrate a dual or limited multi-technology focus for PETs. Korea, for example, focuses on federated learning (30%) and homomorphic encryption (34%), maintaining less significant capacities in other PETs. Concurrently, the United States, the United Kingdom and Israel concentrate around one-third of their portfolios on TEEs. France, meanwhile, directs almost 60% of its capacity on homomorphic encryption and TEEs, with anonymisation having a share of 17%.
Highly specialised portfolios: a small group of the top 15 countries concentrate their inventive efforts in one or two PETs. Finland and Sweden exhibit the most pronounced specialisation, with 49% and 47% of their PET patents related to federated learning, respectively. China, meanwhile, displays a dual concentration in federated learning (41%) and TEEs (28%).

Overall, PET innovation across the top patenting economies is strongly concentrated in a small set of PETs (federated learning, TEEs, anonymisation and homomorphic encryption).

Limited awareness, unclear return on investment and regulatory uncertainty remain primary barriers to PETs adoption

Copy link to Limited awareness, unclear return on investment and regulatory uncertainty remain primary barriers to PETs adoption

The evidence above shows the growing technical maturity of research in PETs. Nevertheless, other work also highlights that widespread adoption remains constrained by a combination of barriers. These barriers include limited awareness and understanding of PET capabilities among decision makers, as well as organisational and regulatory barriers (OECD, 2023_[1]; 2025_[2]; Information Commissioner’s Office (ICO), 2024_[32]). Combined, the evidence and findings underscore the need for concerted intervention from policymakers to address the gap presented by these trends.

Consultations, including by the UK Information Commissioner’s Office (2024_[32]), confirm persistent gaps in understanding PETs and inconsistent terminology. Many organisations remain uncertain about what qualifies as a PET; how different tools function; which ones are ready for deployment; and which have already been deployed in other contexts and sectors. This results in both overestimation of PET capabilities and underestimation of their practicality, with organisations dismissing them as too complex, costly or immature.

The shortage of specialised technical expertise exacerbates these challenges. Deploying PETs requires advanced knowledge of cryptography, distributed systems and privacy engineering, skills that are high in demand, specialised and expensive. Public sector organisations and SMEs often lack in-house capacity to assess or implement these tools, creating dependencies on external consultants and vendors. This bottleneck is particularly acute in healthcare and government – sectors where privacy requirements are among the highest, but where technical expertise is the most constrained.

The lack of awareness and technical expertise leads to organisations struggling to articulate a clear business case for PETs. Because the benefits are largely preventive (reducing the risk and cost of data breaches rather than generating direct revenue), their return on investment is difficult to quantify. Complex pricing models and the absence of standardised metrics for measuring privacy risk reduction further obscure cost-benefit analyses.

Regulatory uncertainty amplifies these obstacles. Although a growing number of data protection and privacy regulators have expressed support for PETs through a variety of initiatives, organisations remain unsure about the actual cost-benefit analysis of using PETs for compliance with existing privacy and data protection frameworks. This uncertainty is further compounded by possible diverging regulatory interpretations, both across sectors (e.g. between data protection and financial market requirements on the fight against money laundering and the financing of terrorism) and across jurisdictions (where divergent thresholds for anonymisation or acceptable risk are applied). Without consistent guidance and recognition that PETs yield benefits in cross-border compliance efforts in an effective manner, organisations may tend to adopt conservative compliance strategies that restrict experimentation and slow PET adoption.

These barriers weigh most heavily on start-ups and specialised PET service providers, which face long sales cycles and demanding customer education processes (see Box 2.S.2). During the Singapore-OECD Expert Workshop on PETs and AI (July 2024), participants described lengthy proof-of-concept demonstrations and pilot projects lasting 12-18 months as challenges for PET service providers. This extended sales cycle can create severe cash flow challenges for early-stage companies that lack the capital reserves of established technology firms. Furthermore, the need for customised solutions for each client (given varying data architectures, regulatory requirements and use cases) may contribute to smaller profit and limited growth potential.

Box 2.S.2. Market dynamics and competition challenges: The case of Inpher

Copy link to Box 2.S.2. Market dynamics and competition challenges: The case of Inpher

Founded in 2015 with backing from investors including JP Morgan and Swisscom Ventures, Inpher developed sophisticated multi-party computation (MPC) protocols, enabling financial institutions and healthcare organisations to perform joint analytics without sharing the underlying data.

Despite raising over USD 25 million and establishing partnerships with major enterprises, Inpher struggled with common PET commercialisation challenges: lengthy sales cycles (often 12-18 months) due to customer education requirements; high implementation costs requiring substantial professional services revenue to offset product licences; and difficulty scaling beyond proof-of-concept deployments to production-grade implementations. The company’s MPC platform, while cryptographically sophisticated, may have required significant computational resources and network bandwidth, limiting its applicability to high-value use cases that could not justify the performance overhead.

The competitive dynamics in the PET market further complicated Inpher’s position. As highlighted by participants at the three OECD Expert Workshops on PETs, small and medium-sized PET service providers face significant challenges when competing against big technology companies, particularly those offering PETs as integrated components of their cloud services. Bundled offerings from established providers are often easier for customers to understand and adopt, as they come with familiar interfaces, existing support infrastructures and the perceived stability of large corporations. Consequently, innovative start-ups with potentially superior technical solutions may be overlooked in favour of less optimal but more accessible offerings from incumbent technology giants.

Moreover, the risk associated with start-up service providers creates an additional barrier. Potential enterprise clients, particularly in regulated industries such as healthcare and finance, typically express concerns about vendor continuity, i.e. the possibility that a start-up might cease operations, leaving critical services unsupported. This creates a negative cycle where customer hesitation limits revenue growth, which in turn increases the actual risk of business failure, ultimately validating the initial concerns.

In 2024, Arcium acquired Inpher primarily for its technical team and intellectual property rather than its customer base. This reflects a broader pattern in the PET industry where technological capabilities are valued but sustainable business models remain elusive. This outcome, despite strong technical foundations and blue-chip investors, underscores that PET commercialisation must go beyond cryptographic innovation. It also requires fundamental advances in performance optimisation, user experience design and market positioning that can compete effectively against both technical limitations and market incumbents.

Source: OECD Expert Workshop on PETs and AI; Swisscom Ventures (2021_[33])Inpher secures strategic investment from Swisscom Ventures, https://ventures.swisscom.com/inpher-secures-strategic-investment-from-swisscom-ventures/; Arcium (2025_[34]), Arcium acquires core tech and team from Inpher, supercharging confidential computing, https://www.arcium.com/articles/arcium-acquires-core-tech-and-team-from-inpher-supercharging-confidential-computing.

Governments and regulators deploy co-ordinated complementary measures to accelerate adoption of PETs

Copy link to Governments and regulators deploy co-ordinated complementary measures to accelerate adoption of PETs

PETs are increasingly addressed, explicitly or implicitly, in countries’ privacy and data protection frameworks. This often takes the form of (i) legal requirements for privacy and data protection by design and by default; (ii) provisions on de-identification, digital security and accountability; and/or (iii) mandates to privacy enforcement authorities to promote adoption of PETs (OECD, 2023_[1]).

To complement these measures, governments and regulators are issuing guidance to clarify compliance pathways. In June 2023, the UK ICO published guidance on how PETs can help organisations comply with UK data protection law (Information Commissioner’s Office (ICO), 2023_[35]), followed by its Guide on Anonymisation (Information Commissioner’s Office (ICO), 2025_[36]). Similarly, the Singapore Personal Data Protection Commission issued a Proposed Guide on Synthetic Data Generation (PDPC, 2024_[37]), while the Israel Privacy Protection Authority released a PET guide in early 2025 (Privacy Protection Authority, 2025_[38]) and a Professional Guide for Medical Data Anonymisation, which includes guidance on the implementation of PETs in the health sector (Privacy Protection Authority; Ministry of Health; Ministry of Justice, 2026_[39]).

Beyond regulatory guidance, governments promote adoption of PETs through demand- and supply-side initiatives. Most target demand creation, encouraging experimentation and uptake, particularly in AI contexts, through a variety of measures and initiatives:

Regulatory sandboxes enable innovators to test PET-based solutions under regulatory supervision. In Singapore, the IMDA/PDPC PET Sandbox allows companies to trial PETs with AI technologies, producing case studies and best-practice guidance (Infocomm Media Development Authority (IMDA), 2024_[40]). Norway’s Datatilsynet Sandbox promotes data protection by design in AI systems (Datatilsynet, 2025_[41]). Meanwhile, the UK ICO Sandbox supports privacy-preserving innovation across sectors, and demonstrates how PETs can be integrated into complex systems as part of a wider data governance framework (ICO, 2025_[42]). The key areas of focus include “Personalised AI” and “Next-generation IoT” (Information Commissioner’s Office (ICO), 2026_[43]), which use PETs to embed privacy from the outset (ICO, 2022_[44]; 2024_[45]).
Innovation contests incentivise creative PET solutions. The UK-US prize challenges (2022-2023) invited innovators to use PETs to combat money laundering and improve pandemic forecasting. Winning projects combined homomorphic encryption, secure MPC and differential privacy (UK-US prize challenges, 2025_[46]). Since 2016, the CNIL-Inria Privacy Award has similarly promoted PET-related research excellence (CNIL, 2026_[47]).
R&D support remains essential to address technical challenges of PETs. The United States National Strategy for Privacy-Preserving Data Sharing and Analytics (NSTC, 2023_[48]) and the Privacy Enhancing Technology Research Act (Stevens, 2024_[49]) commit federal agencies to fund R&D in PETs. Türkiye’s TÜBİTAK provides over 50 grant initiatives supporting PET-related innovation.
Collaboration initiatives link public and private expertise. In Israel, the Ministry of Health collaborates with IBM and start-ups on AI diagnostics using de-identification based on differential privacy. Meanwhile, the UK’s OpenSAFELY enables secure analysis of 55 million patient records through remote computation within protected environments – an approach directly aligned with PET principles.
Government procurement can accelerate diffusion by rewarding PET-integrated solutions. Estonia’s “Bürokratt”, launched in 2022, is the world’s first AI-based virtual public-service assistant using federated learning and synthetic data to safeguard citizen privacy (Information System Authority (RIA), 2025_[50]).
Cost-benefit awareness tools, such as those provided by the UK Department for Science, Innovation and Technology (2024_[51]), can help organisations evaluate potential returns from PET adoption, supporting investment decisions and addressing perceived cost barriers.

Annex 2.S.A. Glossary and typology of privacy-enhancing technologies

Copy link to Annex 2.S.A. Glossary and typology of privacy-enhancing technologies

Privacy-enhancing technologies (PETs) encompass a diverse array of technical approaches and tools, each designed to address various privacy and confidentiality challenges in data processing and sharing. The OECD (2023_[1]) report on emerging PETs categorises them into three main groups based on their functional characteristics, and data governance and privacy objectives (data accountability tools are not covered in this report).

Data obfuscation tools increase privacy protections by altering the data, by adding “noise” or by removing identifying details

Copy link to Data obfuscation tools increase privacy protections by altering the data, by adding “noise” or by removing identifying details

In addition to anonymisation and pseudonymisation tools, data obfuscation tools include the following:

Synthetic data generation creates artificial datasets that preserve statistical properties of original data without containing actual individual records, enabling AI model training and testing without privacy risks (particularly when combined with differential privacy guarantees).
Differential privacy adds calibrated statistical noise to data or computational outputs to prevent individual re-identification, while preserving aggregate patterns. However, determining optimal privacy budgets (epsilon values) remains challenging, particularly for unstructured or multimodal data.

PETs that encrypt data processing allow data to remain encrypted while in use (in-use encryption), avoiding the need to decrypt the data before processing

Copy link to PETs that encrypt data processing allow data to remain encrypted while in use (in-use encryption), avoiding the need to decrypt the data before processing

Homomorphic encryption allows computations to be performed directly on encrypted data without decryption. This enables secure cloud computing and privacy-preserving analytics, although with computational overhead. However, such overhead is becoming less significant with increased computing capacity.
Secure multi-party computation enables multiple parties to jointly compute functions over their combined data without revealing individual inputs, using techniques such as secret sharing and garbled circuits. However, this requires substantial communication bandwidth.
Trusted execution environments provide hardware-based secure enclaves for confidential computing, leveraging processor-level isolation to protect data during processing, with applications ranging from cloud computing to edge devices.

Federated and distributed analytics allows executing analytical tasks upon data that are not visible or accessible to those executing the tasks

Copy link to Federated and distributed analytics allows executing analytical tasks upon data that are not visible or accessible to those executing the tasks

Federated learning trains machine-learning models across distributed datasets without transferring raw data, aggregating only model updates rather than underlying information. This requires co-ordination mechanisms and protection against inference attacks.
Private set intersection identifies common elements across datasets held by different parties without revealing non-matching records, enabling applications such as contact tracing and fraud detection while preserving privacy.

Data accountability tools aim to enhance privacy and data protection by enabling data subjects to control their own data

Copy link to Data accountability tools aim to enhance privacy and data protection by enabling data subjects to control their own data

Data accountability tools are often not considered as PETs in the strict sense given that they do not primarily aim to protect the confidentiality of personal data at a technical level. Most tools are in their early stages of development; have narrow sets of use cases; and lack stand-alone applications. In fact, the number of publications, OSS repositories and patent applications were negligible compared to other types of PETs. That said, further analysis is needed to ensure that the keywords are most appropriate to identify data accountability tools.

Annex 2.S.B. Methodology

Copy link to Annex 2.S.B. Methodology

The analytical framework builds on OECD approaches to tracking emerging technologies through patent and publication data, notably the OECD report Identifying emerging AI technologies using patent data (OECD, 2025_[27]), in line with the guidance from the OECD Patent Statistics Manual (OECD, 2009_[26]). It adapts these methods to PETs using a more simplified approach (excluding, for example, a formal Delphi study for keyword validation).

The keyword taxonomy was derived from the OECD report on emerging privacy-enhancing technologies (PETs) (OECD, 2023_[1]), and on PETs and AI (OECD, 2025_[2]). It covers the four main families of PETs defined and adopted in these reports: (i) data obfuscation; (ii) encrypted data processing; (iii) distributed data processing; and (iv) accountable systems (see Annex 2.S.A).

Supplementary terms were identified through text mining of recent arXiv publications (2020-2025), extracting and filtering unigrams, bigrams and trigrams co-occurring with core PET terms. This led to a list of 20 keywords related to differential privacy, homomorphic encryption, federated learning and trusted execution environments, as well as around 150 alternative terms (See Table 2.S.B.1.). This final taxonomy was used across all three datasets featured in the analysis.

Three complementary datasets were analysed to capture the innovation cycle of PETs. Extensive filtering removed non-technical or duplicate entries to reduce false positives. This led to the following records retrieved and filtered:

arXiv (publications): 30 971 records retrieved; 30 248 created between 2010-2025.
GitHub (open source): 46 581 repositories identified; 9 410 verified after excluding forks and inactive projects; 9 119 created between 2010-2025.
PATSTAT (patents): from the EPO Spring 2025 release, 80 737 (broad definition) and 31 084 (narrow definition) applications were identified. “Narrow” and “broad” definitions of PETs combined keyword and CPC/IPC classification searches as presented in Tables 2.S.B.2 and 2.S.B.3. This analysis uses the broad definition unless otherwise specified. Of the patents identified using the broad definition, 62 714 were filed between 2010-2022. This included 25 857 international families, 23 081 of which were filed between 2010-2022.

Sectoral and AI-related linkages were detected through keyword co-occurrence, with AI-related keywords adopted from the OECD report Identifying emerging AI technologies using patent data (OECD, 2025_[27]). See keywords in Tables 2.S.B.4 and 2.S.B.5.

Inventor country contributions follow OECD fractional counting (see endnote 7), and patent families were de-duplicated via PATSTAT’s DOCDB identifiers. Coverage extends to February 2026 for arXiv/GitHub and 2022 for PATSTAT, due to limitations associated with an 18-month publication lag (see endnote 2).

To complete the incomplete years for PATSTAT (2023-2025), counts for these years were forecasted using a basic AR(1) regression model using annual data. This approach models the value of the dependent variable in the current year as a linear function of its value in the previous year, plus an estimated constant term capturing the average drift. This function is used to forecast missing values using latest available baseline.

Overall, key constraints include incomplete coverage of non-English or proprietary sources, cross-country differences in software and cryptography patentability, recent-year under-reporting and evolving PET terminology (see endnote 2).

Primary term	Alternative terms
privacy enhancing technologies	privacy-enhancing technologies; privacy enhancing technology; privacy-enhancing technology; privacy enhancing; privacy-enhancing; privacy tech; privacy tools; privacy preserving; privacy preserving technologies; privacy preserving technology; confidentiality-enhancing technologies; confidentiality enhancing technologies; confidentiality-enhancing technology; confidentiality enhancing technology; confidentiality enhancing; confidentiality-enhancing; confidentiality-preserving technologies; confidentiality preserving technologies; confidentiality preserving technology; privacy preserving data sharing; privacy preserving data processing
privacy-preserving machine learning	confidential inference; privacy-preserving inference; homomorphic inference; private inference
privacy by design	data protection by design; data minimization; data minimisation by design
data obfuscation	data obfuscation tools; obfuscation tools; de-identification
anonymisation	anonymization; data anonymisation; data anonymization; anonymous data; k-anonymity; k anonymity
pseudonymisation	pseudonymization; data pseudonymisation; pseudonymised information
synthetic data	synthetic data generation; synthetic dataset; data synthesis
differential privacy	-differential privacy; noise addition; noise injection; privacy budget; epsilon differential privacy
zero-knowledge proofs	zero-knowledge; zero knowledge; zero-knowledge proof; zero knowledge proofs; zk-SNARKs; zk-STARKs; ZK proofs
selective disclosure	attribute suppression; column suppression; record suppression; row suppression; character masking; data masking
encrypted data processing	aggregation with encryption
homomorphic encryption	fully homomorphic encryption; somewhat homomorphic encryption; partially homomorphic encryption
multi-party computation	multi party computation; secure multi party computation; secure multiparty computation; multiparty computing
private set intersection	secure set intersection; private matching
trusted execution environments	trusted execution environment; secure enclave; confidential computing
distributed data processing	distributed data analysis; distributed analysis; privacy-preserving distributed; distributed analytics
federated learning	federated training; on-device learning; on-device training; on-device processing; on-device analytics; federated data processing; federated transfer learning; federated analytics
threshold secret sharing	secret sharing; computation threshold signing; Shamir’s secret sharing
accountable systems	privacy accountability
personal data stores	personal data store; personal information management system; data pods; Solid pods

Note: Terms in bold are names of the four PET categories (see Annex 2.S.A). In total ~150 alternative terms were attributed to 20 primary terms.

CPC classification code	Code description
G06F21/6245	Protecting personal data (e.g. for financial or medical purposes)
G06F21/6254	Anonymising data (e.g. decorrelating personal data from owner’s identification)
G06N 3/098	Distributed learning (e.g. federated learning)
H04L9/008	Homomorphic encryption
H04L9/085	Secret sharing or secret splitting (e.g. threshold schemes)

IPC/CPC classification code	Code description
G06F21/60	Privacy computing
G06F21/62	Data protection
G06F21/71	Secure computing or processing of information
G06F21/57	Certifying or maintaining trusted computer platforms
G06F21/53	Executing in restricted environment (e.g. sandbox or secure virtual machine)
H04W12/02	Privacy protection in wireless communication
H04L9/00	Arrangements for secret or secure communication
H04L9/32	Cryptographic mechanisms for secret or secure communication
H04L63/04	Confidential data exchange among entities through data packet networks
G06F16/256	Federated or virtual databases

Domain	Keywords used for co-occurrence
Healthcare & Medical	healthcare; health care; medical data; clinical data; patient data; patient privacy; hospital; electronic health record; ehr; medical image; clinical trial; biomedical; pharmaceutical; drug discovery; genomic data; genome privacy; epidemiological; covid-19; pandemic; telemedicine; radiology; diagnostic; wearable device; medical device
Finance & Banking	financial data; financial institution; bank; banking; payment system; financial transaction; credit card; fraud detection; anti-money laundering; trading; stock market; cryptocurrency; blockchain; loan; insurance; fintech
Industrial & Manufacturing	industrial; manufacturing; industry 4.0; industrial iot; iiot; supply chain; logistics; predictive maintenance; quality control; smart factory; production line; industrial control
E-commerce & Advertising	e-commerce; ecommerce; online shopping; online retail; online marketplace; shopping platform; retail platform; online store; online purchase; recommendation system; recommender system; collaborative filtering; personalized recommendation; content recommendation; product recommendation; personalization; advertising; advertisement; targeted advertising; online advertising; ad auction; ad targeting; marketing analytics; programmatic advertising; behavioral advertising; ad network
Mobility & Transportation	autonomous vehicle; self-driving; connected vehicle; vehicular network; vehicle-to-vehicle; v2v; v2x; automotive; location privacy; location data; location-based service; trajectory privacy; mobility data; gps; geolocation; spatial data
Wireless & Mobile networks	wireless network; 5g; 6g; cellular network; mobile network; base station; wireless communication; mobile communication; spectrum sharing
Social Media & Social Networks	social media; social network; facebook; twitter; instagram; online social; social graph; social computing
Cloud Computing	cloud computing; cloud service; cloud server; cloud storage; multi-cloud; cloud provider; serverless
Education & E-Learning	education; e-learning; online learning; educational data; student data; learning analytics; mooc
Smart Grid & Energy	smart grid; power grid; energy consumption; electricity consumption; smart meter; demand response; renewable energy system
Internet of Things (IoT)	internet of things; iot; smart home; smart city; smart device; sensor network; edge computing; fog computing
Web & Browser	web browser; web tracking; browser; cookie; web privacy; online tracking; tracking protection; fingerprinting
Databases & Data Management	database; data management; query processing; sql; nosql; data warehouse; data lake; postgresql; mysql; mongodb

Keywords used for AI co-occurrence
action recognition activity recognition adaboost adaptive boosting adversarial learning adversarial network ai algorithm ai chipset algorithmic of ai ambient intelligence answer set program ant colony ant colony optimisation artificial bee colony algorithm artificial evolution artificial intelligence artificial neural network association rule attention transformer attention based autoencoder automated planning automated scheduling autonomic computing autonomous car autonomous vehicle autonomous weapon backpropagation bayes network Bayesian learning bayesian network bee colony blind signal separation bootstrap aggregation brain computer interface brownboost classification tree cluster analysis cognitive automation cognitive computing cognitive insight system cognitive modelling cognitive robotic collaborative filtering collision avoidance community detection computational intelligence computational pathology computer vision connectionis convolutional neural convolutional neural network decision model decision tree deep belief network deep convolutional neural network deep learning deep learning model deep neural deep neural network deep reinforcement learning dictionary learning differential evolution algorithm diffusion model dimensionality reduction driverless car driverless vehicle dynamic time warping embedding emotion recognition ensemble learning evolutionary algorithm evolutionary computation evolutionary robotic expert system extreme machine learning face recognition facial expression recognition factorisation machine feature engineering feature extraction feature learning feature selection federated learning firefly algorithm fuzzy c fuzzy environment fuzzy logic fuzzy number fuzzy set fuzzy system gaussian mixture model gaussian process generation model generative adversarial network generative ai genetic algorithm genetic programming gesture recognition gradient boosting gradient tree boosting graph convolutional network graph neural graph neural network graphical model gravitational search algorithm hebbian learning hidden markov model hierarchical clustering high dimensional data high dimensional feature high dimensional input high dimensional model Hough transformer human action recognition human activity recognition human aware artificial intelligence human robot interaction humanoid robot hybrid ai image classification image processing image recognition image retrieval image segmentation independent component analysis inductive logic inductive logic programm inductive monitoring instance based learning intelligence augmentation intelligent agent intelligent classifier intelligent geometric computing intelligent infrastructure intelligent software agent intelligent system intuitionistic fuzzy set K means Kernel learning language transformer language model large language large language model latent dirichlet allocation latent semantic analysis latent variable layered control system learning transformer learning algorithm learning automata learning model link prediction llm logistic regression logitboost long short term memory lpboost machine intelligence machine learning machine learning model machine translation machine vision madaboost MapReduce Markovian memetic algorithm meta learning model training motion planning
multi agent system multi label classification multi layer perceptron multi objective evolutionary algorithm multi sensor fusion multi task learning multilayer perceptron multi objective optimisation natural gradient multinomial naïve Bayes natural language natural language generation naïve Bayes classifier natural language processing nearest neighbour algorithm neural architecture search neural machine translation neural network natural language understanding neural network cnn neural network model neural radiance field neural turing neural turing machine neuromorphic computing neuromorphic architecture neuromorphic engineering non negative matrix factorisation obstacle avoidance one class learning one shot learning object detection object recognition optical character recognition optimization for learning particle swarm optimisation pattern recognition optimisation for learning pedestrian detection policy gradient method Q learning random field random forest planning algorithm recommender system rankboost recurrent neural network regression tree reinforcement learning representation learning relational learning robot interaction robot learning rough set rule learning rule based learning self driving car self organising map self organising structure semi supervised training semi supervised learning sensor data fusion sensor fusion sentiment analysis similarity learning service robot simultaneous localisation mapping single linkage clustering sparse representation speech recognition speech synthesis speech to text spectral clustering stacked generalisation stochastic gradient stochastic gradient descent supervised learning supervised training statistical relational learning support vector machine support vector regression swarm intelligence swarm optimization t s fuzzy system Takagi Sugeno fuzzy system temporal difference learning text mining text to speech topic model totalboost training data trajectory planning trajectory tracking transfer learning trust region policy optimisation unmanned aerial vehicle unsupervised learning unsupervised training variational inference vector machine virtual assistant visual servoing xgboost

Keywords used for AI co-occurrence

action recognition activity recognition adaboost adaptive boosting adversarial learning adversarial network ai algorithm ai chipset algorithmic of ai ambient intelligence answer set program ant colony ant colony optimisation artificial bee colony algorithm artificial evolution artificial intelligence artificial neural network association rule attention transformer attention based autoencoder automated planning automated scheduling autonomic computing autonomous car autonomous vehicle autonomous weapon backpropagation bayes network Bayesian learning bayesian network bee colony blind signal separation bootstrap aggregation brain computer interface brownboost classification tree cluster analysis cognitive automation cognitive computing cognitive insight system cognitive modelling cognitive robotic collaborative filtering collision avoidance community detection computational intelligence computational pathology computer vision connectionis convolutional neural convolutional neural network decision model decision tree deep belief network deep convolutional neural network deep learning deep learning model deep neural deep neural network deep reinforcement learning dictionary learning differential evolution algorithm diffusion model dimensionality reduction driverless car driverless vehicle dynamic time warping embedding emotion recognition ensemble learning evolutionary algorithm evolutionary computation evolutionary robotic expert system extreme machine learning face recognition facial expression recognition factorisation machine feature engineering feature extraction feature learning feature selection federated learning firefly algorithm fuzzy c fuzzy environment fuzzy logic fuzzy number fuzzy set fuzzy system gaussian mixture model gaussian process generation model generative adversarial network generative ai genetic algorithm genetic programming gesture recognition gradient boosting gradient tree boosting graph convolutional network graph neural graph neural network graphical model gravitational search algorithm hebbian learning hidden markov model hierarchical clustering high dimensional data high dimensional feature high dimensional input high dimensional model Hough transformer human action recognition human activity recognition human aware artificial intelligence human robot interaction humanoid robot hybrid ai image classification image processing image recognition image retrieval image segmentation independent component analysis inductive logic inductive logic programm inductive monitoring instance based learning intelligence augmentation intelligent agent intelligent classifier intelligent geometric computing intelligent infrastructure intelligent software agent intelligent system intuitionistic fuzzy set K means Kernel learning language transformer language model large language large language model latent dirichlet allocation latent semantic analysis latent variable layered control system learning transformer learning algorithm learning automata learning model link prediction llm logistic regression logitboost long short term memory lpboost machine intelligence machine learning machine learning model machine translation machine vision madaboost MapReduce Markovian memetic algorithm meta learning model training motion planning

multi agent system multi label classification multi layer perceptron multi objective evolutionary algorithm multi sensor fusion multi task learning multilayer perceptron multi objective optimisation natural gradient multinomial naïve Bayes natural language natural language generation naïve Bayes classifier natural language processing nearest neighbour algorithm neural architecture search neural machine translation neural network natural language understanding neural network cnn neural network model neural radiance field neural turing neural turing machine neuromorphic computing neuromorphic architecture neuromorphic engineering non negative matrix factorisation obstacle avoidance one class learning one shot learning object detection object recognition optical character recognition optimization for learning particle swarm optimisation pattern recognition optimisation for learning pedestrian detection policy gradient method Q learning random field random forest planning algorithm recommender system rankboost recurrent neural network regression tree reinforcement learning representation learning relational learning robot interaction robot learning rough set rule learning rule based learning self driving car self organising map self organising structure semi supervised training semi supervised learning sensor data fusion sensor fusion sentiment analysis similarity learning service robot simultaneous localisation mapping single linkage clustering sparse representation speech recognition speech synthesis speech to text spectral clustering stacked generalisation stochastic gradient stochastic gradient descent supervised learning supervised training statistical relational learning support vector machine support vector regression swarm intelligence swarm optimization t s fuzzy system Takagi Sugeno fuzzy system temporal difference learning text mining text to speech topic model totalboost training data trajectory planning trajectory tracking transfer learning trust region policy optimisation unmanned aerial vehicle unsupervised learning unsupervised training variational inference vector machine virtual assistant visual servoing xgboost

Source: OECD adopted from OECD (2025_[27]), Identifying emerging AI technologies using patent data: A semi-automated approach, OECD Publishing, Paris, https://doi.org/10.1787/d17e9a1a-en.

References

[52] Agre, P. and M. Rotenberg (eds.) (1998), Technology and Privacy: The New Landscape, MIT Press.

[54] Ahn, J. et al. (2023), “Federated Learning for Predictive Maintenance and Anomaly Detection Using Time Series Data Distribution Shifts in Manufacturing Processes”, Sensors (Basel, Switzerland), Vol. 23/17, p. 7331, https://doi.org/10.3390/S23177331.

[34] Arcium (2025), “Arcium acquires core tech and team from Inpher, supercharging confidential computing”, https://www.arcium.com/articles/arcium-acquires-core-tech-and-team-from-inpher-supercharging-confidential-computing (accessed on 15 October 2025).

[23] Ben-Sasson, E. et al. (2014), “Zerocash: Decentralized anonymous payments from bitcoin”, Proceedings – IEEE Symposium on Security and Privacy, pp. 459-474, https://doi.org/10.1109/SP.2014.36.

[24] BIS Innovation Hub (2023), “Project Aurora: The power of data, technology and collaboration to combat money laundering across institutions and borders”, https://www.bis.org/about/bisih/topics/fmis/aurora.htm (accessed on 13 October 2025).

[8] Breschi, S. et al. (2019), “Public research and innovative entrepreneurship: Preliminary cross-country evidence from micro data”, OECD Science, Technology and Industry Policy Papers, No. 64, OECD Publishing, Paris, https://doi.org/10.1787/0d057da7-en.

[53] Byrd, D. and A. Polychroniadou (2020), “Differentially Private Secure Multi-Party Computation for Federated Learning in Financial Applications”, ICAIF 2020 - 1st ACM International Conference on AI in Finance, Vol. 20, https://doi.org/10.1145/3383455.3422562.

[20] Canidio, A. (2018), “Financial incentives for open source development: The case of blockchain”, MPRA Paper, Munich Personal RePEc Archive, https://mpra.ub.uni-muenchen.de/93748.

[47] CNIL (2026), “Launch of 10th edition of CNIL-Inria privacy award”, https://www.cnil.fr/en/launch-10th-edition-cnil-inria-privacy-award (accessed on 10 March 2026).

[22] ConsenSys (2026), “Blockchain for digital identity | Real world blockchain use cases”, https://consensys.net/blockchain-use-cases/digital-identity (accessed on 30 January 2026).

[14] Cyberspace Administration of China (2016), 中华人民共和国网络安全法 [Cybersecurity Law of the People’s Republic of China], https://www.cac.gov.cn/2025-12/29/c_1768735112911946.htm (accessed on 9 March 2026).

[41] Datatilsynet (2025), “Regulatory privacy sandbox”, https://www.datatilsynet.no/en/regulations-and-tools/sandbox-for-artificial-intelligence (accessed on 18 March 2025).

[19] Durkee, M. et al. (2025), “Using privacy enhancing technologies to enable international data sharing”, Blog Government Digital Service, https://gds.blog.gov.uk/2025/10/09/using-privacy-enhancing-technologies-to-enable-international-data-sharing (accessed on 11 October 2025).

[12] European Union (2016), “Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data, and repealing Directive 95/46/EC (General Data Protection Regulation)”, No. 119/1, EUR-Lex, https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX:32016R0679 (accessed on 30 January 2026).

[30] Fink, C. et al. (2022), Resilience and Ingenuity: Global Innovation Responses to Covid-19, Centre for Economic and Policy Research, Paris, https://cepr.org/publications/books-and-reports/resilience-and-ingenuity-global-innovation-responses-covid-19.

[10] Gentry, C. (2009), “Fully homomorphic encryption using ideal lattices”, CMU School of Computer Science, https://www.cs.cmu.edu/~odonnell/hits09/gentry-homomorphic-encryption.pdf.

[56] Hard, A. et al. (2018), Federated Learning for Mobile Keyboard Prediction, https://research.google/pubs/pub47586 (accessed on 30 January 2026).

[5] IBM (2025), Cost of a Data Breach Report 2025, IBM, Armonk, NY, https://www.ibm.com/reports/data-breach (accessed 30 January 2026).

[42] ICO (2025), Regulatory Sandbox Final Report: Meta Platforms, Inc, Information Commissioner’s Office, London, https://ico.org.uk/media2/bfcjqc32/meta-regulatory-sandbox-final-report.pdf.

[44] ICO (2022), Tech Horizons Report, Information Commissioner’s Office, London, https://ico.org.uk/media2/about-the-ico/documents/4023338/ico-future-tech-report-20221214.pdf.

[40] Infocomm Media Development Authority (IMDA) (2024), “Privacy enhancing technology sandboxes”, https://www.imda.gov.sg/how-we-can-help/data-innovation/privacy-enhancing-technology-sandboxes (accessed on 18 March 2025).

[36] Information Commissioner’s Office (ICO) (2025), “Anonymisation”, https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/data-sharing/anonymisation (accessed on 10 June 2025).

[45] Information Commissioner’s Office (ICO) (2024), “Personalised AI”, https://ico.org.uk/about-the-ico/research-reports-impact-and-evaluation/research-and-reports/technology-and-innovation/tech-horizons-and-ico-tech-futures/tech-horizons-report-2024/personalised-ai (accessed on 29 January 2026).

[32] Information Commissioner’s Office (ICO) (2024), Tackling Barriers to Privacy-Enhancing Technologies Adoption – a PETs Project Report, Information Commissioner’s Office, London, https://ico.org.uk/about-the-ico/research-reports-impact-and-evaluation/research-and-reports/technology-and-innovation/tackling-barriers-to-privacy-enhancing-technologies-adoption.

[43] Information Commissioner’s Office (ICO) (2026), “Our current areas of focus for the regulatory sandbox”, https://ico.org.uk/for-organisations/advice-and-services/regulatory-sandbox/our-key-areas-of-focus-for-the-regulatory-sandbox (accessed on 30 January 2026).

[35] Information Commissioner’s Office (ICO) (2023), “Privacy-enhancing technologies (PETs)”, https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/data-sharing/privacy-enhancing-technologies (accessed on 30 January 2026).

[50] Information System Authority (RIA) (2025), “Bürokratt”, https://www.ria.ee/en/state-information-system/personal-services/burokratt (accessed on 30 January 2026).

[25] Mastercard (2024), “Stepping through the looking glass of privacy-enhancing technologies”, https://www.mastercard.com/us/en/news-and-trends/stories/2024/stepping-through-the-looking-glass-of-privacy-enhancing-technologies.html (accessed on 30 January 2026).

[9] McMahan, H. et al. (2016), “Communication-efficient learning of deep networks from decentralized data”, Proceedings of the 20th International Conference on Artificial Intelligence and Statistics, AISTATS 2017, https://doi.org/10.48550/arXiv.1602.05629.

[11] Microsoft (2026), “Microsoft SEAL”, https://www.microsoft.com/en-us/research/project/microsoft-seal (accessed on 30 January 2026).

[16] National People’s Congress of the People’s Republic of China (2021), Data Security Law of the People’s Republic of China, http://www.npc.gov.cn/englishnpc/c2759/c23934/202112/t20211209_385109.html (accessed on 9 March 2026).

[15] National People’s Congress of the People’s Republic of China (2021), Personal Information Protection Law of the People’s Republic of China, http://en.npc.gov.cn.cdurl.cn/2021-12/29/c_694559_2.htm (accessed on 9 March 2026).

[48] NSTC (2023), National Strategy to Advance Privacy-preserving Data Sharing and Analytics, National Science and Technology Council, Washington, DC, https://www.nitrd.gov/pubs/National-Strategy-to-Advance-Privacy-Preserving-Data-Sharing-and-Analytics.pdf (accessed on 30 January 2026).

[18] OECD (2026), “Trends in AI incidents and hazards reported by the media”, OECD Artificial Intelligence Papers, No. 53, OECD Publishing, Paris, https://doi.org/10.1787/4f5ff43c-en.

[3] OECD (2025), Enhancing Access to and Sharing of Data in the Age of Artificial Intelligence, OECD Publishing, Paris, https://doi.org/10.1787/23a70dca-en.

[21] OECD (2025), “Facilitating the secondary use of health data for public interest purposes across borders”, OECD Digital Economy Papers, No. 376, OECD Publishing, Paris, https://doi.org/10.1787/d7b90d15-en.

[27] OECD (2025), Identifying emerging AI technologies using patent data: A semi-automated approach, OECD Publishing, Paris, https://doi.org/10.1787/d17e9a1a-en.

[4] OECD (2025), “Mapping relevant data collection mechanisms for AI training”, OECD Artificial Intelligence Papers, No. 48, OECD Publishing, Paris, https://doi.org/10.1787/3264cd4c-en.

[2] OECD (2025), “Sharing trustworthy AI models with privacy-enhancing technologies”, OECD Artificial Intelligence Papers, No. 38, OECD Publishing, Paris, https://doi.org/10.1787/a266160b-en.

[55] OECD (2024), “Defining AI incidents and related terms”, OECD Artificial Intelligence Papers, OECD Publishing, Paris, https://doi.org/10.1787/d1a8d965-en.

[1] OECD (2023), “Emerging privacy-enhancing technologies: Current regulatory and policy approaches”, OECD Digital Economy Papers, No. 351, OECD Publishing, Paris, https://doi.org/10.1787/bf121be4-en.

[7] OECD (2020), “The public sector innovation lifecycle: A device to assist teams and organisations in developing a more sophisticated approach to public sector innovation”, OECD Working Papers on Public Governance, No. 37, OECD Publishing, Paris, https://doi.org/10.1787/0d1bf7e7-en.

[6] OECD (2015), Frascati Manual 2015: Guidelines for Collecting and Reporting Data on Research and Experimental Development, The Measurement of Scientific, Technological and Innovation Activities, OECD Publishing, Paris, https://doi.org/10.1787/9789264239012-en.

[26] OECD (2009), OECD Patent Statistics Manual, OECD Publishing, Paris, https://doi.org/10.1787/9789264056442-en.

[17] OECD-GPAI (2026), “AI incidents and hazards monitor”, https://oecd.ai/en/incidents (accessed on 28 September 2026).

[37] PDPC (2024), Privacy Enhancing Technology (PET) Proposed Guide on Synthetic Data Generation, Personal Data Protection Commission, Singapore.

[38] Privacy Protection Authority (2025), Guide To Privacy Enhancing Technologies (PETs), Privacy Protection Authority, Israel, https://www.gov.il/BlobFolder/reports/guide_enhancing_technologies/en/PETs-Mongash-en.pdf.

[39] Privacy Protection Authority; Ministry of Health; Ministry of Justice (2026), מדריך מקצועי להתממת מידע רפואי [Professional guide for medical data anonymisation], Privacy Protection Authority, Ministry of Health, Ministry of Justice, Israel, https://www.gov.il/BlobFolder/generalpage/medinfoguide/he/media_media_MedInfoGuide2026.pdf.

[13] State of California (2018), California Consumer Privacy Act, State of California, US, https://leginfo.legislature.ca.gov/faces/codes_displayText.xhtml?division=3.&part=4.&lawCode=CIV&title=1.81.5 (accessed 30 January 2026).

[49] Stevens, H. (2024), H.R.4755 - 118th Congress (2023-2024): Privacy Enhancing Technology Research Act, https://www.congress.gov/bill/118th-congress/house-bill/4755 (accessed 18 March 2025).

[33] Swisscom Ventures (2021), “Inpher secures strategic investment from Swisscom Ventures”, https://ventures.swisscom.com/inpher-secures-strategic-investment-from-swisscom-ventures (accessed on 15 October 2025).

[31] Trunschke, M. et al. (2024), “Pandemic effects: Do innovation activities of firms suffer from long COVID?”, Research Policy, Vol. 53/7, p. 105024, https://doi.org/10.1016/J.RESPOL.2024.105024.

[51] UK Department for Science, Innovation and Technology (2024), “Cost-benefit awareness tool”, https://www.gov.uk/government/publications/privacy-enhancing-technologies-cost-benefit-awareness-tool/cost-benefit-awareness-tool (accessed on 30 October 2025).

[46] UK-US prize challenges (2025), “Privacy-enhancing technologies”, https://petsprizechallenges.com (accessed on 18 March 2025).

[57] USPTO (2021), Trademarks and Patents in China: The Impact of Non-Market Factors on Filing Trends and IP Systems, United States Patent and Trademark Office, Washington, DC, https://www.uspto.gov/sites/default/files/documents/USPTO-TrademarkPatentsInChina.pdf.

[58] WIPO (2024), World Intellectual Property Indicators 2024, World International Property Organization, Geneva, https://www.wipo.int/edocs/pubdocs/en/wipo-pub-941-2024-en-world-intellectual-property-indicators-2024.pdf.

[29] WIPO (2023), Global Innovation Tracker, World Intellectual Property Organization, Geneva, https://www.wipo.int/edocs/pubdocs/en/wipo-pub-2000-2023-section1-en-global-innovation-tracker-global-innovation-index-2023.pdf.

[28] WIPO (2022), World Intellectual Property Report 2022: The Direction of Innovation, World International Property Organization, Geneva, https://doi.org/10.34667/tind.45356.

Notes

Copy link to Notes

← 1. The three expert workshops in question are: the United Kingdom-Estonia-OECD Expert Workshop on PETs (May 2024); the Singapore-OECD Expert Workshop on PETs and AI (July 2024); and the Canada-Japan-OECD Expert Workshop and High-Level Roundtable on PETs and AI held alongside the G7 Data Protection Authorities’ Meeting in Ottawa (June 2025).

← 2. This paper is based on the analysis of 30 248 arXiv papers, 9 119 GitHub repositories and 62 714 PATSTAT patent applications, filed or published between 2010 and 2025 (except for PATSTAT data, where 2022 was the cut-off date). The latest extraction for each data source occurred in February 2026. A few limitations need to be acknowledged given the selection of data sources. First, while arXiv offers timely access to cutting-edge research outputs, it does not cover the full scientific literature (e.g. it excludes proprietary or peer-reviewed databases such as SCOPUS). Therefore, it may underrepresent some disciplines and research publications. Similarly, GitHub represents the largest OSS platform, yet not the entire OSS ecosystem, and many projects are hosted on other repositories. Third, patent data reflect only one route from concept to commercialisation. As emphasised in OECD (2025_[27]), patents do not protect all inventions and patentability can vary significantly by jurisdiction. In particular, national patent systems differ in whether and how they recognise software and cryptographic methods, i.e. computer-implemented or algorithmic inventions, which may lead to cross-country comparability issues. See Annex S.2.B for further details on the methodology.

← 3. GitHub stars enable users to bookmark repositories as a signal of interest or usefulness.

← 4. Academic publications (i.e. arXiv) and OSS repositories (i.e. GitHub repositories) capture different stages of the innovation pathway for PETs. Academic publications primarily reflect the development and diffusion of the underlying scientific and technical knowledge base, whereas OSS repositories provide usable tools, libraries and applications that practitioners can adopt. Given the open-source nature of PET development, OSS activity serves as a useful proxy for early-stage implementation and experimentation outside formal commercial deployment. Consequently, a comparison of their trajectories provides indicative evidence of whether implementation activity is expanding faster or slower than foundational research.

← 5. The OECD defines an event where the development or use of an AI system results in actual harm as an “AI incident”, while an event where the development or use of an AI system is potentially harmful is termed an “AI hazard” (OECD, 2024_[55]).

← 6. As the large majority of papers and repositories at the intersection of Finance & Banking and Zero Knowledge are blockchain related, this intersection serves as a suitable proxy for blockchain activity.

← 7. International patent families, meaning inventions that have been filed in more than one jurisdiction or through the Patent Cooperation Treaty (PCT), are typically associated with higher-value or more commercially significant innovations, since international protection entails higher costs. To avoid double counting across countries, each patent family is fractionally attributed to the inventors’ countries of residence. For example, a patent listing two inventors in Japan and one in Germany would allocate two-thirds to Japan and one-third to Germany. This fractional counting approach provides a more accurate measure of each country’s actual contribution to global inventive activity, regardless of where a company files or owns the patent (OECD, 2009_[26]).

← 8. The data used to calculate the number of ICT-related patent families are provisional and do not yet capture all relevant IPC codes. As a result, the absolute shares presented here may slightly overestimate the proportion of ICT patent families accounted for by PET patents. However, there is no reason to expect this limitation to have significant bearing on growth of this share over time.

← 9. The data used to calculate these growth rates are based on the narrow definition of PETs. The OECD average growth rate is calculated as the growth rate of the total number of all fractional counts attributed to OECD Member countries between 2020 and 2022.

← 10. China’s apparent dominance in patent volume must be interpreted with caution due to several confounding factors. According to WIPO (2024_[58]), applicants from China accounted for around 68% of all global patent families in 2021 and roughly 69% of global patent filings in 2023. Moreover, approximately 97% of patent families originating from China during 2019-2021 were single-office filings (i.e. limited international extension). These shares suggest that the high volume of filings from China may reflect strong domestic filing incentives and quantity-driven institutions rather than equivalent levels of globally oriented, multi-jurisdictional innovation. Additionally, USPTO (2021_[57]) highlights many government subsidy programmes, performance-linked targets for patents in universities and state-owned enterprises, and other non-market drivers of filing behaviour in China. All these factors caution against interpreting patent counts alone as a direct measure of underlying innovation quality.

Featured topics

Agriculture and fisheries

Climate change

Development

Digital

Economy

Education and skills

Employment

Environment

Finance and investment

Governance

Health

Industry, business and entrepreneurship

Regional, rural and urban development

Science, technology and innovation

Society

Taxation

Trade

Energy

Nuclear energy

Transport

Featured topics

Agriculture and fisheries

Climate change

Development

Digital

Economy

Education and skills

Employment

Environment

Finance and investment

Governance

Health

Industry, business and entrepreneurship

Regional, rural and urban development

Science, technology and innovation

Society

Taxation

Trade

Energy

Nuclear energy

Transport

Countries A - C

Countries D - I

Countries J - M

Countries N - R

Countries S - T

Countries U - Z

Regional and global engagement

Countries

Countries A - C

Countries D - I

Countries J - M

Countries N - R

Countries S - T

Countries U - Z

Regional and global engagement

Publications

Publications

Featured publications

Data

Data

Featured data

News & events

News & events

Featured events

About OECD

About

Engage with us

Work with us

Featured topics

Agriculture and fisheries

Climate change

Development

Digital

Economy

Education and skills

Employment

Environment

Finance and investment