PrivDNN: A Secure Multi-Party Computation Framework for Deep Learning using Partial DNN Encryption

Researchers introduced PrivDNN, a privacy-preserving deep learning framework for secure DNN inference.

It combines secure multi-party computation, partial model encryption, and core-neuron selection to protect both model privacy and user data privacy.

Tests show it can significantly reduce inference time and memory use compared with heavier encrypted-model approaches.

The result makes secure AI evaluation more practical for machine learning as a service and other private inference settings.