Federated Learning, Synthetic Data, and Other PETs

Federated Learning, Synthetic Data, and Other PETs: A Complete CDPSE Guide

Introduction
Privacy-Enhancing Technologies (PETs) are tools and techniques that let organizations collect, process, analyze and share data while reducing exposure of personal information. In the ISACA CDPSE (Certified Data Privacy Solutions Engineer) exam, PETs sit mainly within the Privacy Architecture and Data Lifecycle domains. They are where privacy-by-design principles become concrete engineering decisions. This guide covers federated learning, synthetic data, and the wider PET family, and then shows how to answer exam questions about them.

Why PETs Are Important
1. Regulatory compliance: Laws such as GDPR (Article 25, data protection by design and by default; Article 32, security of processing), CCPA/CPRA, HIPAA and LGPD expect organizations to use appropriate technical measures. PETs are the clearest example of such measures.
2. Data minimization: PETs reduce the amount of identifiable data collected, moved or exposed. This directly supports the minimization and purpose-limitation principles.
3. Unlocking data value: PETs make analytics, AI/ML training, research and cross-organization collaboration possible without centralizing raw personal data.
4. Risk reduction: Less exposure of personal data means a smaller breach impact, a smaller attack surface and lower regulatory penalties.
5. Cross-border transfers: Some PETs allow insights to move across borders while the raw data stays local, which supports data localization requirements.
6. Trust: Customers, partners and regulators increasingly expect organizations to show privacy-respecting innovation.

What Are PETs? Core Categories
PETs can be grouped by what they protect and how:
- Data-shifting or decentralizing PETs: federated learning, edge processing, on-device analytics.
- Data-altering PETs: synthetic data, differential privacy, anonymization, pseudonymization, tokenization, masking, generalization, k-anonymity, l-diversity and t-closeness.
- Computation-protecting (cryptographic) PETs: homomorphic encryption, secure multi-party computation (SMPC), zero-knowledge proofs, private set intersection.
- Hardware-based PETs: Trusted Execution Environments (TEEs) and confidential computing, such as Intel SGX, AMD SEV and ARM TrustZone.

1. Federated Learning (FL)
What it is: A machine learning approach in which a model is trained across many decentralized devices or servers that each hold local data. The raw data is never sent to a central location. Only model updates are shared, such as gradients or weights.
How it works:
Step 1: A central server sends an initial global model to participating clients (phones, hospitals, branches).
Step 2: Each client trains the model locally on its own data.
Step 3: Clients send only their model updates back to the server.
Step 4: The server aggregates the updates, often with Federated Averaging (FedAvg), to produce an improved global model.
Step 5: The cycle repeats until the model converges.
Variants:
- Horizontal FL: same features, different users. Example: mobile keyboards.
- Vertical FL: same users, different features. Example: a bank and a retailer.
- Cross-device vs. cross-silo: many small devices vs. a few large institutions.
Privacy benefits: Data minimization, data localization and reduced centralized breach risk.
Limitations and risks: Model updates can still leak information. Known attacks include gradient inversion, membership inference and model poisoning. FL is not a complete privacy guarantee on its own. It is usually combined with secure aggregation (often based on SMPC), differential privacy (noise added to updates) and TEEs. It also brings communication overhead, problems with non-identically distributed data, and the challenge of governing the participants.
Use cases: Predictive text on smartphones, healthcare models trained across hospitals, fraud detection across banks.

2. Synthetic Data
What it is: Artificially generated data that keeps the statistical properties and patterns of a real dataset but does not contain actual records of real individuals.
How it works:
Step 1: A generative model learns the distributions and correlations in the source data. Common approaches are statistical models, Bayesian networks, GANs (Generative Adversarial Networks), VAEs (Variational Autoencoders) and large language models.
Step 2: The model samples new artificial records.
Step 3: The output is validated for utility (does it behave like the real data?) and for privacy (could any record be traced back to a real person?).
Types:
- Fully synthetic: no real records at all.
- Partially synthetic: only sensitive attributes are replaced.
- Hybrid: a mix of real and synthetic records.
Privacy benefits: Supports test and development environments, data sharing with vendors or researchers, AI training and demos without exposing real personal data.
Limitations and risks:
- Overfitted generators can memorize and reproduce real records, which creates re-identification risk. Outliers are especially vulnerable.
- Utility can be lost, and rare cases may be underrepresented.
- Bias in the source data can be reproduced or amplified.
- Synthetic data generated with differential privacy gives stronger, measurable guarantees.
- Synthetic data is not automatically outside the scope of privacy law. If individuals can be re-identified, it may still count as personal data.

3. Differential Privacy (DP)
A mathematical framework that adds calibrated noise to data or query results. The result is that including or excluding any single individual barely changes the output. Privacy loss is measured by the parameter epsilon (the privacy budget). A smaller epsilon means stronger privacy but less accuracy.
- Central (global) DP: a trusted curator adds the noise.
- Local DP: noise is added on the user's device before collection.
Examples include the US Census Bureau, Apple and Google. The main trade-off is privacy vs. utility, and the privacy budget gets used up by repeated queries.

4. Homomorphic Encryption (HE)
Allows computation directly on encrypted data. The result stays encrypted and, once decrypted, matches the result of computing on the plaintext.
- Partially HE: supports one operation.
- Somewhat HE: supports limited operations.
- Fully HE: supports arbitrary computation.
Typical use is outsourcing computation to an untrusted cloud. The main drawback is heavy computational overhead.

5. Secure Multi-Party Computation (SMPC/MPC)
Several parties jointly compute a function over their combined inputs while each keeps its own input private, often using secret sharing. Examples include joint salary benchmarking, fraud analytics across banks, and the secure aggregation step in FL. It requires significant communication and coordination between parties.

6. Zero-Knowledge Proofs (ZKP)
One party proves that a statement is true without revealing the underlying data. Examples are proving age over 18 without revealing a birth date, and privacy-preserving authentication or identity verification.

7. Trusted Execution Environments (TEEs) / Confidential Computing
Hardware-isolated secure enclaves protect data in use, even from the host operating system or cloud provider. They rely on trust in the hardware vendor and can be vulnerable to side-channel attacks.

8. Private Set Intersection (PSI)
Two parties find the records they have in common without revealing the records they do not share. Examples are contact discovery and ad-conversion measurement.

9. Traditional De-identification PETs
- Pseudonymization: replaces identifiers with pseudonyms. It is reversible with additional information, so under GDPR the data is still personal data.
- Anonymization: irreversible. If done properly, the data falls outside GDPR.
- Tokenization: substitutes sensitive values with tokens, using a vault for mapping. Common for PCI data.
- Data masking: static or dynamic obfuscation, often used in non-production environments.
- Generalization and suppression: k-anonymity, l-diversity and t-closeness.

How to Select the Right PET (Decision Logic)
- Data cannot leave its location (residency, sensitivity, many data holders) → Federated learning or edge processing.
- Need realistic data for testing, development, sharing or AI training without real individuals → Synthetic data, ideally with DP.
- Need aggregate statistics or public releases with mathematically measurable protection → Differential privacy.
- Need an untrusted party to compute on data it must never see → Homomorphic encryption.
- Several organizations want a joint result without sharing inputs → SMPC (or PSI for finding overlaps).
- Need to prove a claim without disclosing the data behind it → Zero-knowledge proof.
- Need to protect data in use within cloud infrastructure → TEE / confidential computing.
- Need re-linkage capability internally while reducing exposure → Pseudonymization / tokenization.

Governance Considerations for Privacy Engineers
- PETs support privacy programs but do not replace them. A DPIA, a lawful basis, transparency and accountability are still required.
- Assess residual re-identification risk regularly, especially as new auxiliary datasets become available.
- Document the PET selection rationale, its parameters (for example, epsilon) and validation results.
- Consider maturity, performance cost, scalability, skills needed and vendor dependency.
- Use layered (defense-in-depth) combinations, such as FL + secure aggregation + DP.
- Align with standards and guidance, for example ISO/IEC 27559 (de-identification), ISO/IEC 20889, NIST Privacy Framework, NIST SP 800-188, ENISA guidance, and ICO guidance on PETs.

Exam Tips: Answering Questions on Federated Learning, Synthetic Data, and Other PETs
1. Identify the core problem in the scenario. Ask yourself: Is the issue data movement, data sharing, computation by an untrusted party, statistical release, test data, or proof of an attribute? Match the problem to the PET using the decision logic above.
2. Keyword mapping:
- "Data never leaves the device or hospital" or "train a model across locations" → federated learning.
- "Realistic test data", "no real customer records", "share with developers or vendors" → synthetic data.
- "Noise", "privacy budget", "epsilon", "aggregate statistics" → differential privacy.
- "Compute on encrypted data", "cloud processes without decrypting" → homomorphic encryption.
- "Multiple parties, joint computation, no one sees others' inputs" → SMPC.
- "Prove without revealing" → zero-knowledge proof.
- "Protect data in use", "secure enclave" → TEE.
3. Remember the residual-risk caveats. ISACA often tests whether you know that no PET is a silver bullet. FL updates can leak data, synthetic data can memorize records, and pseudonymized data is still personal data. When an answer claims a PET "eliminates all privacy risk," it is usually wrong.
4. Prefer layered or combined solutions. If the options include "federated learning combined with differential privacy and secure aggregation" versus FL alone, the combined option is typically the better answer.
5. Think like a privacy engineer and manager. Before choosing a PET, the BEST first step is often to perform a privacy impact assessment or data classification, or to understand the processing purpose. Pick the governance-first answer when the question asks what to do FIRST.
6. Know the trade-offs.
- DP: privacy vs. accuracy.
- HE: security vs. performance.
- Synthetic data: utility vs. privacy.
- FL: decentralization vs. communication cost and leakage via gradients.
Questions may ask for the PRIMARY disadvantage or limitation.
7. Distinguish anonymization from pseudonymization. This is a classic exam trap. Reversible means pseudonymized, which is still regulated. Irreversible means anonymized, which may fall outside scope. Properly generated synthetic data may approach anonymization, but it must be validated.
8. Non-production environments: When a question asks how to protect personal data in testing or development, the best answers are synthetic data or data masking, not copying production data with access controls.
9. Watch qualifiers: BEST, MOST, PRIMARY, FIRST, GREATEST risk. Several answers may be technically valid. Choose the one that most directly addresses the stated objective with the least exposure of personal data, in line with data minimization.
10. Validation questions: For synthetic data, expect references to testing re-identification risk (membership inference, nearest-neighbor distance) and statistical fidelity. For DP, expect questions on setting and tracking the privacy budget.
11. Legal and regulatory linkage: PETs demonstrate privacy by design (GDPR Article 25) and appropriate technical measures (Article 32). If asked how an organization can demonstrate compliance through engineering, PETs plus documentation is the answer.
12. Eliminate distractors. Encryption at rest or in transit does not protect data in use. Access controls do not provide data minimization. Hashing alone is weak pseudonymization because identifiers with small value spaces can be brute-forced.

Sample Practice Questions
Q1: Several hospitals want to build a diagnostic AI model, but regulations prohibit moving patient records outside each hospital. Which approach is MOST appropriate?
Answer: Federated learning, ideally with secure aggregation and differential privacy, because the raw data stays on-site.

Q2: A development team needs realistic customer data for application testing. What is the BEST privacy-preserving option?
Answer: Generate synthetic data validated for re-identification risk, rather than copying production data.

Q3: What is the PRIMARY residual privacy risk of federated learning?
Answer: Inference of training data from shared model updates, through gradient inversion or membership inference.

Q4: An organization wants a cloud provider to perform calculations on sensitive financial data without the provider ever seeing the plaintext. Which PET fits BEST?
Answer: Homomorphic encryption. A TEE or confidential computing is an alternative if hardware trust is acceptable.

Q5: Increasing the epsilon value in differential privacy will:
Answer: Increase accuracy and utility but decrease privacy protection.

Summary
Federated learning moves the model to the data. Synthetic data replaces real data with statistically similar artificial data. Other PETs either add noise (DP), encrypt computation (HE, SMPC), prove claims without disclosure (ZKP), or isolate processing in hardware (TEE). For the CDPSE exam:
- Map each scenario's objective to the right PET.
- Recognize that every PET has limitations and residual risk.
- Favor layered approaches.
- Remember that PETs operate within a broader governance framework of DPIAs, lawful basis, documentation and continuous risk assessment.