Privacy-Enhancing Technologies (PETs)

• PETs are technical tools, protocols, and organizational measures that safeguard digital privacy by obfuscating sensitive information while enabling collaborative computation.
• They include diverse methods such as differential privacy, secure multi-party computation, homomorphic encryption, and federated learning, each with unique trade-offs in performance and utility.
• Deployment of PETs in regulated sectors supports compliance, risk management, and innovative applications in AI and IoT, demanding careful balancing of privacy and data utility.

Privacy-Enhancing Technologies (PETs) are a class of technical tools, protocols, and organizational mechanisms designed to mitigate privacy risks inherent in digital information processing ecosystems. PETs operationalize privacy by enabling collaborative computation, storage, and analysis of data while ensuring that sensitive or personal information is obfuscated, minimized, partitioned, or otherwise protected against unauthorized inference or exposure. In current deployments, PETs are central to legal compliance, risk management, and the sociotechnical negotiation of trust in regulated and competitive environments.

1. Foundations and Conceptual Taxonomies

PETs are systematically categorized by their technical approach, privacy goals, and integration context. Comprehensive taxonomies integrate dimensions including scenario (e.g., client/server trust configuration), privacy aspect (identity, content, or behavior), aims (indistinguishability, unlinkability, deniability, confidentiality), underlying technical foundations (e.g., cryptographic security model), relevant data lifecycle phase (storage, transmission, processing), reliance on trusted third parties, and reversibility (Shafieinejad et al., 4 Mar 2025). For practical architectural selection, dimensions such as technology maturity, performance impact, architectural overhead, and intrinsic data utility degradation are crucial (Kunz et al., 2022). The socio-technical framing, grounded in regulatory, technical, and organizational interplay, positions PET adoption as a dynamic negotiation among actors responding to legal mandates, technological advancements, and market/cultural expectations (Shafieinejad et al., 4 Mar 2025).

PET maturity within industry can be bracketed into three tiers:

• Common but non-SOTA: basic access control, k-anonymization for data release.
• Mature and widespread: multi-factor authentication, SSL/TLS, AES encryption.
• State-of-the-art (SOTA): differential privacy (DP), secure multi-party computation (MPC), homomorphic encryption (HE), trusted execution environments (TEE), oblivious RAM (ORAM), federated learning, synthetic data (Shafieinejad et al., 4 Mar 2025).

2. Principal PET Families and Technical Mechanisms

PETs are classified by their core privacy principle and technical apparatus. The principal families include:

PET Class	Formal Guarantee/Definition	Representative Primitives
Differential Privacy (DP)	For all $D, D'$ differing in one record and all $S$,$$\Pr[M(D)\in S]\leq e^\varepsilon\Pr[M(D')\in S]+\delta$$	Laplace/Gaussian mechanism, DP-SGD
Secure Multi-Party Computation (MPC)	$n$ parties compute $f(x_1,...,x_n)$; each party learns only $f(x)$	Garbled Circuits, Secret Sharing
Homomorphic Encryption (HE)	Supports computation on ciphertexts: $\Dec(\Eval(f, \Enc(m_1), ..., \Enc(m_t)))=f(m_1,...,m_t)$	CKKS, BGV, RLWE-based FHE
Trusted Execution Environment (TEE)	Hardware-enforced code/data secrecy via isolated enclave	Intel SGX, AWS Nitro, remote attestation
Federated Learning (FL)	Local training, global aggregation: $$\theta_{t+1} = \text{Aggregate}(\{\Delta \theta^i_t\}_i)$$	FedAvg, secure aggregation
Anonymization	$k$-Anonymity, $\ell$-Diversity, $t$-Closeness	Generalization, suppression
Onion/Garlic Routing	Multi-layer encryption, randomized relay paths to break linkability	Tor, I2P, Garlic-Onion Routing (GOR)
Synthetic Data	Statistical mimicry: $P_{\theta} \approx P_{\text{data}}$	GANs, Copulas, VAEs

Each PET imposes a distinct set of trust and threat models, privacy metrics (e.g., indistinguishability advantage $\varepsilon$ for DP, adversary success in cryptographic games), and performance-utility trade-offs (Oualha, 17 Jun 2025, Kunz et al., 2022).

3. Applications, Integration Patterns, and Empirical Practices

PETs are embedded in diverse pipelines, including:

• Secure computation and ML pipelines via hybrid designs (e.g., MPC+HE, FL+DP) to enable privacy-preserving analytics, distributed AI training, and encrypted inference (Sedghighadikolaei et al., 2024, d'Aliberti et al., 2024).
• Privacy-preserving data sharing in regulated sectors (e.g., healthcare GDPR compliance via Multiparty Homomorphic Encryption) (Scheibner et al., 2020).
• End-to-end privacy engineering in Internet of Things (IoT), notably wearables leveraging federated learning, homomorphic encryption, and on-demand blockchain ledgers for fine-grained consent and auditability (Barma et al., 5 Mar 2025, Mosaiyebzadeh et al., 2023).

PET pipeline integration requires careful mapping to architectural stages:

• Data generation/storage: searchable encryption, ORAM.
• Local/edge computation: TEE, lightweight HE.
• Model/data sharing: federated aggregation, secure summation.
• Output/inference: DP perturbation, ZKP-based attestation (Sedghighadikolaei et al., 2024, Barma et al., 5 Mar 2025, Oualha, 17 Jun 2025).

Deployment choices must weigh overhead and formal strength. For example, FHE achieves robust protection at $10^3$–$10^5\times$ computational cost over plaintext, while DP incurs tunable accuracy loss dependent on chosen $\varepsilon$ (privacy budget) (Oualha, 17 Jun 2025, d'Aliberti et al., 2024). Hybrid PETs (e.g. MPC-TEE, HE-MPC, FL+DP) offer composable trade-offs between trust assumptions, performance, and security (Sedghighadikolaei et al., 2024).

4. Power, Governance, and Contextual Risks

Recent empirical analyses detail how PETs can be instrumented to reinforce provider power, rather than mitigating privacy asymmetries. In the Amazon Sidewalk case, classic PETs—end-to-end encryption and identifier obfuscation—are deployed to offer "narrow" privacy assurances to device owners but simultaneously expand Amazon’s infrastructural dominance and telemetry scope. PETs in this mode foster:

• Suppression of some information flows to legitimize new, pervasive gateway deployments (covering >90% of US population), thus enabling novel surveillance and competition risks via location/usage telemetry and cross-device data aggregation.
• Infrastructural lock-in by mandating Amazon-certified hardware, factory key-injection, and AWS-only data decryption—thereby reconfiguring manufacturer production and R&D as accessories to provider infrastructure (Gend et al., 2024).

This "PET Paradox" highlights the insufficiency of equating privacy with technical secrecy; robust power analysis must account for control of device/production environments, mandatory cloud dependency, and emergent information flows. Regulators are urged to adopt privacy-competition-sovereignty metrics, push for device-sovereignty (legal end-user OS control), and support decentralized PET management (Gend et al., 2024).

5. Evaluation Methodologies and Adoption Dynamics

Methodologically rigorous PET selection incorporates:

• Enhanced taxonomies combining privacy protection goal, scenario fit, maturity, performance, architectural impact, and utility loss (Kunz et al., 2022).
• Process models integrating regulatory mapping, utility-cost-benefit analyses, and cross-functional team design (Shafieinejad et al., 4 Mar 2025).
• Lifecycle mapping—need identification, solution exploration, evaluation, approval, and post-approval—reflecting the convergence of legal, technical, and organizational triggers (Shafieinejad et al., 4 Mar 2025).

Empirical adoption in industry is shaped by:

• Compatibility/integration with workflows and existing tech stacks.
• Regulatory triggers and risk management calculus.
• Market and client expectations.
• Organizational resources and cross-team communication (Shafieinejad et al., 4 Mar 2025).

Human factors research indicates that advanced PETs (e.g., Tor, VPN, encryption) are least adopted, with practical barriers in usability, information deficits, perceived necessity, cost, trust, and absence of social reinforcement. Layered roll-out strategies, embedded education, packaging, and quantified feedback are essential for scaling adoption beyond integrated, "simple" PETs (Coopamootoo, 2020, Boteju et al., 2023).

6. Trade-Offs: Utility, Performance, and Environmental Impact

PET deployment in high-assurance and AI-enabled systems presents multidimensional trade-offs:

• Privacy vs. utility: Lower $\varepsilon$ in DP tightens privacy but reduces model accuracy; $k$-anonymity/generalization reduces identifiability but also data granularity (Oualha, 17 Jun 2025, Kunz et al., 2022).
• Performance and resource overhead: FHE and MPC introduce orders-of-magnitude latency/memory penalties compared to plaintext; HE for ML inference/training may induce up to $10^5\times$ energy and carbon footprint over non-private baselines (Damie et al., 6 Aug 2025).
• Energy and carbon: Carbon reporting for cryptographic PETs shows wide variability, from modest ($\sim 1.7\times$ for HTTPS) to prohibitive (100,000$\times$ for homomorphic ML), underscoring the need for performance-aware design and sustainable alternatives (Damie et al., 6 Aug 2025).

7. Limitations, Contextual Integrity, and Future Research

PETs often instantiate privacy as confidentiality and do not, by themselves, guarantee contextually appropriate information flows. Using Nissenbaum’s contextual integrity, deployments in age verification, content scanning, and model training can realize technical guarantees while violating entrenched social or legal norms if repurposed for broader or secondary data flows (Balsa et al., 2023). This highlights the risk that PETs, if narrowly scoped, can be co-opted for privacy-invasive or anticompetitive purposes.

Priority open research domains include:

• Systematic mapping of SOTA PET use versus industry mislabeling.
• Standardized compliance benchmarks connecting PET parameters (e.g., DP $\varepsilon$) to regulatory criteria.
• Design of lightweight, utility-preserving PETs for resource-constrained and IoT environments.
• PET frameworks for external AI audit and governance, leveraging composable enclaves, SMPC, and ZKP (Beers et al., 5 Feb 2025).
• Full life-cycle carbon/energy benchmarking and optimization for sustainability (Damie et al., 6 Aug 2025).

Careful alignment of PET design with contextual, organizational, and social requirements remains essential to ensuring PETs enhance privacy as normatively understood, rather than simply providing technical opacity.