Privacy-Enhancing Technologies

1. Why Privacy-Enhancing Technologies Matter for AML Professionals

Financial crime does not respect institutional or national boundaries. Money launderers deliberately spread activity across multiple banks, payment providers, fintechs and jurisdictions so that no single institution sees the full picture. A bank looking only at its own data may see a few unremarkable transactions, while the complete network, visible only when data from several institutions is combined, reveals a layering scheme or a mule network.

The obvious solution is to share data. The obstacle is privacy and data protection law. Regimes such as the EU General Data Protection Regulation (GDPR), the UK Data Protection Act, bank secrecy laws, and sector-specific confidentiality rules restrict the transfer and processing of personal data. Tipping-off rules and the confidentiality of Suspicious Activity Reports (SARs/STRs) add further limits. For years this created a tension: regulators expected institutions to detect complex cross-institution laundering, yet the law limited the data they could lawfully pool.

Privacy-Enhancing Technologies (PETs) are the bridge across this gap. They allow institutions to gain the analytical benefit of combined data while keeping the underlying personal data confidential, minimised, or never disclosed at all. The Financial Action Task Force (FATF) highlighted their importance in its July 2021 report Stocktake on Data Pooling, Collaborative Analytics and Data Protection, and in its 2021 work on Opportunities and Challenges of New Technologies for AML/CFT. Supervisors such as the UK FCA, the Monetary Authority of Singapore (through its COSMIC platform), and FinCEN have run or endorsed pilots. The EU AML package (AMLR) explicitly contemplates information-sharing partnerships. For the CAMS candidate, PETs sit squarely within the Monitoring, Investigation and AI Technology domain because they enable the collaborative analytics, shared typology detection and AI model training that modern monitoring depends on.

2. What Privacy-Enhancing Technologies Are

A Privacy-Enhancing Technology is any technical method that lets data be analysed, shared or used to generate insight while reducing or eliminating the exposure of personal or confidential information. The common thread is data minimisation: you extract the value of the data without revealing more than necessary.

It is helpful to divide PETs into three broad families:

(a) Techniques that transform or disguise the data
- Anonymisation: irreversibly removing identifiers so that individuals cannot be re-identified. Truly anonymised data generally falls outside data protection law, but genuine anonymisation is hard to achieve because combinations of indirect identifiers (date of birth, postcode, transaction patterns) can allow re-identification.
- Pseudonymisation: replacing identifiers with a token or key. The data can be re-linked to the person by whoever holds the key. Under GDPR, pseudonymised data is still personal data, but pseudonymisation is a recognised safeguard.
- Tokenisation and hashing: converting a value such as an account number or national ID into a fixed token or hash. Two institutions hashing the same identifier with the same method can detect a match without exchanging the raw identifier.
- Differential privacy: adding carefully calibrated statistical noise to query results or datasets so that aggregate patterns remain accurate but no single individual's inclusion can be inferred. Useful for publishing statistics or training models on sensitive data.
- Synthetic data: artificially generated datasets that mimic the statistical properties of real data (for example, realistic transaction flows) without containing real customers. Used to develop, test and train monitoring models without exposing real records.

(b) Techniques that allow computation on data without revealing it
- Homomorphic encryption: a form of encryption that allows calculations to be performed directly on encrypted data; when the result is decrypted it matches the result that would have been obtained on the plaintext. A bank could send encrypted transaction data to a shared analytics engine, which computes a risk score without ever seeing the actual data.
- Secure Multi-Party Computation (SMPC or MPC): a cryptographic protocol in which several parties jointly compute a function over their combined inputs while each party's input stays secret. For example, five banks can determine whether a customer appears at more than two of them without any bank disclosing its customer list.
- Zero-Knowledge Proofs (ZKPs): a method by which one party proves a statement is true (for example, this customer has passed KYC, or this customer is over 18 and not on a sanctions list) without revealing the underlying data that makes it true.
- Trusted Execution Environments (TEEs) / confidential computing: secure hardware enclaves inside a processor where data is decrypted and processed in isolation, so that even the operator of the server or cloud cannot access it.

(c) Techniques that bring the analysis to the data instead of moving the data
- Federated learning: a machine learning approach where the model is trained locally at each institution on its own data; only model parameters or updates (not the raw data) are shared and aggregated into a global model. Each bank benefits from patterns learned across the network, while customer data never leaves the bank.
- Federated or distributed analytics: similar in spirit, queries are pushed to the data holders and only aggregated or minimised results are returned.

3. How PETs Work in Practice Within AML/CFT

Collaborative transaction monitoring. A consortium of banks uses SMPC or homomorphic encryption to run shared detection rules across their collective data. Suppose funds move rapidly from Bank A to Bank B to Bank C. Each bank alone sees one leg; the collaborative platform recognises the full layering chain. Alerts are returned to the relevant institution, which then investigates using its own data and files an STR if warranted. Singapore's COSMIC platform and the UK's various information-sharing pilots are real-world examples of this model.

Shared watchlists and risk indicators. Institutions hash identifiers of known mule accounts or suspected fraud actors. A private set intersection (a form of MPC) lets another institution check whether any of its customers match, without either side revealing its full list. This supports exit-and-re-entry detection, where a customer closed for suspicious activity opens an account elsewhere.

Public-private partnerships (PPPs). Financial Intelligence Units and law enforcement share typologies and sometimes target indicators with banks. PETs allow institutions to test whether those indicators appear in their data and report back in aggregate, without the authority obtaining unfiltered access to customer records.

AI and machine learning model development. Transaction monitoring models trained on one institution's data often lack enough confirmed laundering cases to learn effectively. Federated learning pools the learning without pooling the data, producing stronger models and fewer false positives. Synthetic data lets vendors and model validators test systems without touching real customer information.

KYC and digital identity. Zero-knowledge proofs and verifiable credentials allow a customer to prove that identity verification has been completed by a trusted provider, or that they satisfy an eligibility rule, without transmitting the underlying documents. This supports reliance arrangements and reduces repeated collection of sensitive data.

Cross-border group monitoring. Global banking groups face data localisation rules preventing customer data from leaving certain jurisdictions. Federated approaches and confidential computing allow group-wide monitoring of risk while respecting local restrictions.

4. Benefits and Limitations

Benefits:
- Enables detection of cross-institution and network-based laundering that siloed monitoring misses.
- Supports compliance with data protection principles of data minimisation, purpose limitation and security by design.
- Reduces the legal and reputational risk associated with large data transfers and breaches.
- Improves AI model quality and reduces false positives through broader training data.
- Builds trust between private institutions and with public authorities.

Limitations and risks:
- Computational cost: homomorphic encryption and MPC are far slower than plaintext processing and may not yet scale to real-time monitoring of millions of transactions.
- Not a substitute for legal basis: PETs reduce privacy risk but do not automatically make processing lawful. Institutions still need a legal gateway for sharing, a defined purpose, and governance. FATF stressed that regulators and data protection authorities must clarify these frameworks.
- Re-identification risk: pseudonymised or poorly anonymised data can sometimes be re-identified, especially when combined with other datasets.
- Tipping-off and SAR confidentiality: sharing must be structured so that the existence of an STR is never disclosed to another institution or inferred by the subject.
- Model risk and explainability: federated models can be harder to validate and explain; regulators expect institutions to understand how alerts are generated.
- Governance and liability: participants must agree who owns outputs, who acts on alerts, how disputes are resolved and how accuracy is maintained.
- Concentration and cyber risk: a shared platform becomes a high-value target.

5. Key Regulatory and Guidance Touchpoints to Remember
- FATF 2021: Stocktake on Data Pooling, Collaborative Analytics and Data Protection, encouraging responsible use of PETs and dialogue between AML and data protection authorities.
- FATF 2021: Opportunities and Challenges of New Technologies for AML/CFT.
- GDPR Article 25 (data protection by design and by default) and Article 32 (security), which explicitly reference pseudonymisation.
- EU AML Regulation provisions enabling information-sharing partnerships under supervisory oversight.
- USA PATRIOT Act Section 314(b), the voluntary information-sharing safe harbour between US financial institutions, which PETs can operationalise at scale.
- National pilots: MAS COSMIC (Singapore), UK FCA TechSprints and the Dutch Transaction Monitoring Netherlands initiative.

6. Exam Tips: Answering Questions on Privacy-Enhancing Technologies

Tip 1: Anchor every answer in the core tension. CAMS questions on PETs are almost always testing whether you understand that PETs resolve the conflict between effective AML information sharing and data protection or bank secrecy obligations. If an answer choice describes a technology that allows analysis or sharing without exposing underlying personal data, it is probably the one the examiner wants.

Tip 2: Know the definitions and match the technique to the scenario. Expect scenario-based questions such as: several banks want to detect customers who hold accounts at multiple institutions without disclosing their customer lists. The correct match is secure multi-party computation (or private set intersection). If the scenario is about training a shared machine learning model without moving customer data, the answer is federated learning. If the scenario involves computing on data while it remains encrypted, the answer is homomorphic encryption. If it involves proving a fact without revealing the data behind it, think zero-knowledge proof. If it involves adding statistical noise to protect individuals in aggregate outputs, think differential privacy.

Tip 3: Distinguish anonymisation from pseudonymisation. This is a classic distractor. Anonymisation is irreversible and takes data outside most data protection regimes if done properly; pseudonymisation is reversible with a key and the data remains personal data. An answer claiming pseudonymised data is no longer personal data is wrong.

Tip 4: Remember that PETs are an enabler, not a legal exemption. If an option suggests that using a PET removes the need for a lawful basis, consent, regulatory approval or governance, reject it. The exam favours answers reflecting that technology plus a proper legal and governance framework are both required.

Tip 5: Watch for tipping-off traps. Any arrangement in which one institution learns that another has filed an STR on a shared customer, or in which the customer could infer an investigation, breaches confidentiality rules. The best collaborative models share risk indicators or alerts, not SAR filings.

Tip 6: Link PETs to the broader technology domain. Questions may embed PETs in wider discussions of AI, machine learning, RegTech or public-private partnerships. Be ready to explain that PETs improve AI model effectiveness (more training data, fewer false positives) and support PPPs by enabling two-way sharing within legal limits.

Tip 7: Use the FATF framing. When a question asks about the regulatory view, recall that FATF encourages responsible innovation, data pooling and collaborative analytics, and calls for cooperation between AML supervisors and data protection authorities. Avoid answers that portray regulators as opposed to information sharing or as indifferent to privacy.

Tip 8: Apply the best answer discipline. Several options may be partially true. Choose the one that most directly solves the problem described with the least disclosure of personal data and the strongest alignment with AML objectives. Eliminate options that involve bulk transfer of raw customer data, that ignore cross-border data restrictions, or that rely on a single institution's view when the scenario clearly describes a multi-institution problem.

Tip 9: Memorise the limitations list. Questions may ask what the principal drawback of a given PET is. Homomorphic encryption and MPC: computational intensity and scalability. Anonymisation: re-identification risk and loss of analytical detail. Federated learning: model validation, explainability and the possibility that model updates leak information if poorly designed. Synthetic data: may not capture rare real-world laundering patterns.

Tip 10: Keep the investigator's perspective. Finally, remember why this matters for the investigator: PETs widen the lens. A PET-enabled alert tells an investigator that a customer's activity is part of a broader network, which should elevate the priority of the case, inform the scope of the investigation, and strengthen the narrative of any resulting STR, all while the investigator continues to work only with data their institution is legally entitled to hold.

7. One-Paragraph Summary for Rapid Revision

Privacy-Enhancing Technologies are technical methods such as anonymisation, pseudonymisation, differential privacy, synthetic data, homomorphic encryption, secure multi-party computation, zero-knowledge proofs, trusted execution environments and federated learning that allow institutions to analyse, share and learn from data without exposing the underlying personal information. In AML/CFT they enable collaborative transaction monitoring, shared risk indicators, stronger AI models and public-private partnerships while respecting data protection, bank secrecy and tipping-off rules. FATF and leading supervisors encourage their responsible use, but PETs complement rather than replace a lawful basis, governance and sound model risk management. In the exam, identify the privacy-versus-detection tension, match the specific technique to the scenario, reject options that treat PETs as a legal shortcut, and select the answer that maximises detection with the minimum disclosure of personal data.