Exam AIGP topic 1 question 73 discussion

The best method to ensure it is statistically unlikely that a specific piece of training data can be identified from an AI model’s output is C. Differential privacy. Differential privacy adds controlled noise to the training process or model outputs, ensuring that the inclusion or exclusion of any single data point has a negligible impact on the results. This provides a mathematical guarantee that individual training data cannot be easily inferred, protecting privacy while maintaining model utility. Why not the others? A. Homomorphic encryption: This allows computations on encrypted data without decrypting it, but it’s primarily used for secure computation, not for preventing data identification in model outputs. B. Data sharding: This involves splitting data across multiple servers to limit access, but it doesn’t inherently protect against data inference from model outputs. D. Data compartmentalization: This separates data into isolated subsets, but it doesn’t provide statistical guarantees against identifying training data from model behavior. Differential privacy is specifically designed to address the risk of data leakage through model outputs, making it the most effective choice.