New to Rust? Grab our free Rust for Beginners eBook Get it free →
Risks of PETs: Uncovering the Dark Secrets of Privacy Enhancing Technologies

When a team plans to analyze personal data, it has to limit exposure without losing the result it needs. Privacy-enhancing technologies (PETs) are methods that let data be used or shared while limiting exposure of the underlying information.
This article explains the risks PETs can leave behind and gives you a sequence for assessing a deployment.
TL;DR: PETs reduce specific risks, not every risk
Privacy-enhancing technologies (PETs) can reduce exposure of personal data, but their protection depends on the method, its configuration, its outputs and the people operating it.
- Define the threat model before choosing a PET.
- Check what data leaves each device or organization, including model updates and results.
- Measure privacy and utility together, then audit deployment and governance.
What are privacy-enhancing technologies?
Privacy-enhancing technologies are methods that let data be used, analyzed or shared while limiting exposure of the underlying information. The term describes a family of techniques, not a single guarantee or certification.
Differential privacy adds calibrated randomness to results so an individual record has bounded influence. Federated learning trains a model across separate data stores without centralizing raw examples. Homomorphic encryption supports computation on encrypted values, while secure multiparty computation lets participants calculate a result without revealing inputs to one another.
These methods protect different points in a data flow. Encryption at rest protects stored data, and encryption in transit protects network traffic. A PET may address data during analysis, but it does not automatically secure the devices, access policies, outputs or people around that analysis.
The CodeForGeek guide to privacy-enhancing technologies covers the broader methods and examples.
What risks should you assess?
A PET deployment can fail through weak assumptions, implementation defects, revealing outputs or overconfident claims. The risks depend on the technique and on who can access each stage.
Privacy guarantees can be misunderstood
Differential privacy is not simply “adding noise.” Its guarantee depends on a defined privacy budget, the mechanism used and how repeated queries compose. If a service answers many related questions, the total privacy loss can grow.
Data can leak through updates and outputs
Federated learning keeps source examples on local devices or servers, but that does not mean all information stays local. Model updates can reveal information about training examples under some conditions. NIST describes privacy attacks on federated learning and cautions that federated learning alone is not a complete privacy solution.
Aggregated results can also expose sensitive facts when groups are small or queries can be repeated and compared. Review the complete interface: inputs, intermediate messages, logs, model updates, reports and outputs. A system that protects raw records can still disclose a sensitive conclusion.
Synthetic data may retain identifying signals
Rare combinations, duplicated examples or memorized training data may expose information about individuals represented in the source.
Test whether records can be linked back to people or whether a model can distinguish training examples from new ones. Check the generator, the source data and the intended release. Synthetic data is not a substitute for a documented disclosure-risk assessment.
Cryptography shifts trust rather than removing it
Homomorphic encryption and secure multiparty computation can reduce what participating organizations learn about one another’s inputs. Yet implementations still depend on correct protocols, keys, access rules and result handling. A party may infer sensitive facts from an allowed result, or operational mistakes may expose data outside the protected computation.
A trusted execution environment isolates computation in a protected hardware region. That shifts part of the trust decision to the hardware, firmware, configuration and attestation process. The relevant question is not whether the method sounds secure, but which attacker it resists and what assumptions remain.
Complexity and cost can weaken deployment
Some PETs require specialized expertise, extra computation or coordination among independent parties. The burden varies by method and workload, so claims that all PETs are expensive or energy intensive are too broad. A rushed integration, poorly managed keys or an unreviewed dependency can undermine a sound design.
Operational cost also changes who can adopt a technique. Smaller organizations may lack the staff or infrastructure to run it safely, while well-resourced providers can gain more control over shared data systems. Include maintenance, monitoring and independent review in the deployment estimate, not only initial compute.
Complexity creates human failure modes as well. A privacy budget can be configured incorrectly, keys can be copied into an unprotected environment, or a result can be exported to a system outside the intended boundary.
A design review should name the tasks that require specialist judgment and ensure those tasks have an owner. Documentation should describe safe defaults and tell operators how to detect a misconfiguration before sensitive data is processed.
Privacy claims can create false confidence
A PET is a technical control, not a full privacy program. It does not by itself establish a lawful purpose, limit collection, set retention periods, provide accountability or make a product’s privacy claim accurate. The FTC notes that implementations differ in how much access they remove and that PETs do not replace an organization’s broader privacy obligations.
How to assess the risks of a PET deployment
Review the proposed data flow from source to result, then test the claims against a specific threat model. The sequence below works for a design review, vendor evaluation or internal privacy assessment.
Step 1: Name the data, purpose and adversary
List the sensitive attributes, the intended analysis and the parties who can observe each stage. Include insiders, service providers, other participants and people who receive the final result. Specify whether the concern is direct identification, inference, unauthorized reuse or exposure after a breach.
- What is the protected asset?
- Who are you protecting it from?
- What access or inference would count as a failure?
Step 2: Trace every boundary
Draw where raw data, encrypted data, keys, updates and outputs move. Mark which organization or component can see each item.
For federated learning, include model updates and aggregation. For a secure enclave, include attestation, administrator access and the code loaded into the protected region.
Check whether the deployment also needs ordinary controls such as authentication, encryption in transit, backups, retention limits and incident response. PETs complement those controls rather than replacing them. Keep the controls in the same review.
Step 3: Inspect the guarantee and its assumptions
Ask for the mechanism, threat assumptions and configuration that support the claim. For differential privacy, request the privacy budget and accounting across repeated releases.
For cryptographic protocols, establish which parties must remain independent and how keys are handled. For confidential computing, understand the hardware and attestation assumptions.
Step 4: Test outputs and failure paths
Evaluate whether repeated or narrowly scoped queries can reveal a person, and whether updates or logs expose source information. Test the expected workflow and plausible misuse cases with data that reflects rare groups, not only average records. Record what happens when a participant, service or security check fails.
Step 5: Compare privacy with utility
Measure whether the protected result remains fit for its purpose. Noise, aggregation and restricted access can reduce accuracy or exclude small populations. A weaker privacy setting may improve a metric while increasing disclosure risk, so document who accepts that trade-off and why.
Step 6: Assign owners and review evidence
Identify the person responsible for keys, configuration, access reviews, updates, incident response and explaining the public claim. Ask for independent testing, audit logs and a process for fixing defects. NIST’s PETs Testbed frames evaluation around suitability for a specific context, rather than assuming one tool fits every use.
Common errors when assessing PET risks
A risk review can miss exposure when it treats a PET label as a complete guarantee or checks only the raw input. Use the questions below to catch assumptions that need evidence before deployment.
| Common error | Check instead |
|---|---|
| Assuming federated learning keeps every signal private | Review model updates, aggregation, logs and released outputs |
| Reading a differential privacy claim without its scope | Request the mechanism, budget and accounting across releases |
| Treating synthetic data as anonymous by definition | Test linkage and memorization risk for the intended release |
| Evaluating only the cryptographic method | Review keys, access, implementation and result handling |
What PETs do not guarantee
PETs do not guarantee anonymity, security against every attacker, fairness or legal compliance. Those conclusions require evidence about the whole system, its use and its governance.
A privacy assessment should consider data minimization first: do not collect information that the task does not need. Then combine appropriate PETs with access controls, secure software practices, clear retention rules and oversight. The OECD describes PETs as complementing privacy and confidentiality frameworks, while noting technical, regulatory and adoption challenges.
| Review question | Evidence to request |
|---|---|
| What does the method protect? | Threat model, protected data and explicit exclusions |
| What can leave the protected boundary? | Data-flow map for inputs, updates, logs and outputs |
| What weakens the guarantee? | Configuration, budget, trust assumptions and failure behavior |
| Who maintains it? | Named owners, audit evidence and incident process |
Conclusion: Match the PET to its threat model
PETs can make useful data analysis possible with less exposure, but each method protects a different boundary. Evaluate the guarantee, outputs, operating conditions and governance before treating a privacy claim as evidence of protection.
For a next step, map one proposed use from source data to final output and compare each boundary with the threat it is meant to address. First-party references include the FTC discussion of PET promises, the NIST PETs Testbed and its federated-learning privacy attacks overview.
What are the main risks of privacy-enhancing technologies?
Risks include implementation errors, information leakage through updates or outputs, privacy and utility trade-offs, difficult audits, operational burden and misleading claims. The exact risks depend on the method and deployment.
Is federated learning a privacy-enhancing technology?
Federated learning is commonly treated as a PET because training data can remain distributed rather than being collected centrally. Model updates and outputs can still leak information, so federated learning alone is not a complete privacy solution.
Does differential privacy guarantee that nobody can identify a person?
No. Differential privacy gives a formal bound on how much an individual record can influence an output under stated assumptions. Its protection depends on the mechanism, privacy budget and repeated releases, and it is not a blanket anonymity guarantee.




