Enabling Large-Scale Privacy-Preserving Recurrent Neural Networks with Fully Homomorphic Encryption (G32b)
Fully homomorphic encryption (FHE) can be used to secure a variety of ML models. The speaker will show that FHE can be applied to large-scale recurrent neural networks (RNNs) and used for image classification and speaker identification. This talk will present a ground-breaking, large-scale, privacy-preserving RNN that has multiple layers, >12M parameters, and >130 timesteps, achieving a practical, end-to-end, privacy-preserving speaker identification network. The speaker will also propose a technique that can be applied to any discretized neural network, ensuring almost perfect correctness while encrypted.