CDPSE Enhancing Technologies 4 — Questions and Answers
Question 1: Which technique enables a researcher to determine whether a specific individual's data is in a dataset without the organization revealing the dataset contents?
- Homomorphic encryption
- Private information retrieval (PIR) (Correct answer)
- Differential privacy
- Secure multi-party computation
Correct answer: Private information retrieval (PIR)
Private information retrieval (PIR) allows a user to query a database and retrieve a record without the database owner learning which record was accessed.
Question 2: Format-preserving encryption (FPE) is preferred over standard encryption in some data systems because:
- It provides stronger cryptographic guarantees than AES
- It produces ciphertext in the same format as plaintext, minimizing application changes (Correct answer)
- It does not require key management infrastructure
- It is faster than all other encryption algorithms
Correct answer: It produces ciphertext in the same format as plaintext, minimizing application changes
FPE maintains the format (length, character set) of the original data, allowing encrypted values to pass validation rules and minimizing changes to downstream systems.
Question 3: In the context of PETs, what is 'data vault tokenization' primarily designed to protect against?
- Unauthorized changes to data at rest
- Exposure of sensitive data in case of a breach by storing tokens instead of real values in application systems (Correct answer)
- SQL injection attacks on the token vault database
- Insider threats from database administrators
Correct answer: Exposure of sensitive data in case of a breach by storing tokens instead of real values in application systems
Data vault tokenization ensures that even if the application database is breached, attackers only obtain tokens with no value outside the secure vault, limiting breach impact.
Question 4: Which emerging privacy-enhancing technology uses cryptographic commitments to allow auditors to verify data processing compliance without accessing the underlying data?
- Differential privacy
- Zero-knowledge proofs (Correct answer)
- Synthetic data
- K-anonymity
Correct answer: Zero-knowledge proofs
Zero-knowledge proofs allow a party to prove that a computation was performed correctly or that data meets certain criteria without revealing the actual data to the verifier.
Question 5: A financial services firm wants to analyze transaction patterns across multiple banks without any bank revealing its customers' individual transactions to the others. The most appropriate PET is:
- Differential privacy applied independently at each bank
- Secure multi-party computation (SMPC) (Correct answer)
- Pseudonymization with a shared key
- Federated learning on aggregated statistics
Correct answer: Secure multi-party computation (SMPC)
SMPC allows multiple banks to jointly compute analytics on their combined data without any party ever seeing another party's raw transaction records.
Question 6: The 't-closeness' model extends l-diversity by requiring that:
- At least t distinct sensitive values exist across the full dataset
- The distribution of sensitive attributes within each group closely matches the distribution in the overall dataset (Correct answer)
- Each record appears in at least t different equivalence classes
- The anonymization algorithm runs in t or fewer processing steps
Correct answer: The distribution of sensitive attributes within each group closely matches the distribution in the overall dataset
T-closeness requires the distribution of sensitive values in any equivalence class to be no further than a threshold t from the overall dataset distribution, preventing skewness attacks.
Question 7: Which of the following is the MOST significant privacy challenge when deploying large language models (LLMs) trained on personal data?
- LLMs cannot comply with GDPR because they process natural language
- LLMs may memorize and reproduce training data containing personal information upon prompting (Correct answer)
- LLMs always require consent from all individuals whose data was used
- LLMs must use differential privacy, making them too inaccurate for practical use
Correct answer: LLMs may memorize and reproduce training data containing personal information upon prompting
Research has demonstrated that LLMs can memorize verbatim sequences from training data, meaning personal information could be extracted through targeted prompts.
CDPSE Enhancing Technologies 4
Which technique enables a researcher to determine whether a specific individual's data is in a dataset without the organization revealing the dataset contents?