Selective encryption speeds up private data AI inference substantially
Encrypt What Matters: When Selective Homomorphic Inference Is Efficient
Cryptography and SecurityMachine Learning
Summary
Performing AI computations on encrypted data keeps information private but is very slow when encrypting everything. The authors study only encrypting small, sensitive parts of the input data while processing the rest openly, which maintains the same results as full encryption without needing to retrain models. They find that AI architectures that keep data processing local see much bigger speed improvements from this selective approach, while ones with broad early mixing don’t gain much. This work shows that how AI models handle data locally is key to efficient private computation.
What this means in practice
- •For cloud service providers: Use selective encryption to speed up privacy-preserving AI predictions on sensitive parts of customer data while reducing computation costs.
- •For mobile app developers: Implement privacy-sensitive features that process only encrypted sensitive input regions efficiently on-device or remotely without full data encryption overhead.
Authors
Ali Backour, Juan Reyes, Jaime Punyed, Ana Onoprishvili
Abstract
Fully homomorphic encryption (FHE) enables inference on private data without revealing it to the server, but evaluating an entire input under FHE is expensive. We study \emph{selective homomorphic inference}, where only a sensitive region of interest (ROI) is encrypted, and computations independent of that region are performed in plaintext. Selective evaluation produces the same output as full FHE on the same model, without retraining. Its efficiency depends on how quickly encrypted dependencies spread through the network. For small encrypted ROIs, locality-preserving architectures can achieve order-of-magnitude homomorphic-evaluation speedups, whereas architectures with early global mixing provide essentially no speedup. These results identify locality as the key architectural property governing the benefit of selective homomorphic inference.