Which of the following is NOT a privacy-preserving ML technique listed?

• A. Differential privacy
• B. Data anonymization
• C. Federated learning
• D. Secure multiparty computation

Answer: (B)

In privacy-preserving ML, the goal is to protect individuals’ data while still enabling useful learning. The techniques that provide formal or practical privacy protections include adding carefully calibrated noise to outputs or data (differential privacy), keeping raw data on devices and sharing only model updates to reduce data exposure (federated learning), and enabling joint computations without revealing inputs (secure multiparty computation). Data anonymization, while it removes direct identifiers, often leaves behind indirect identifiers and patterns that can be linked back to individuals, especially when combined with other data sources. It also lacks formal guarantees against re-identification in real-world scenarios and with high-dimensional data. Because of these weaknesses, data anonymization does not meet the same robust privacy standards as the other techniques, making it the one that does not fit as a privacy-preserving ML technique in this list.