Which statement best describes privacy-preserving inference?

• A. Data scraping from public web pages.
• B. Visualizing model predictions.
• C. Training models on distributed datasets with no security.
• D. Techniques that allow performing inference on encrypted or isolated data without exposing raw data (e.g., homomorphic encryption, secure enclaves).

Answer: (D)

Privacy-preserving inference is about getting predictions from a trained model without exposing the sensitive data used to produce those predictions. This is achieved by performing computations on encrypted data or inside protected hardware environments so that raw inputs (and sometimes outputs) remain confidential. Techniques like homomorphic encryption allow operations on ciphertexts, while secure enclaves or trusted execution environments keep data isolated during processing. This is crucial when handling sensitive information (health, finance, personal data) and you want to use models without revealing the underlying data.

The other statements don’t describe this protective approach: scraping data from public pages doesn’t involve keeping data private during inference; visualizing model predictions is simply about showing results; training on distributed data without security pertains to the data's protection during training, not inference.