Inference Attacks

Inference attacks use outputs, models, embeddings, statistics, or auxiliary data to learn sensitive facts.

Examples

• Reconstruction attacks from gradients or statistics
• Attribute inference from released aggregates
• Embedding inversion from vector search systems
• Model extraction and prompt leakage

Why PETs Still Need Output Review

Protecting inputs during computation does not guarantee that the released output is safe.

PET Implications

Use DP, thresholds, output review, query limits, model auditing, and release documentation.