RHODE: Privacy-preserving federated RNN training and prediction based on multi-party homomorphic encryption
Privacy-Preserving Federated Recurrent Neural Networks
AI summary
RHODE enables RNNs to perform privacy-preserving training and prediction in cross-silo federated learning using multi-party homomorphic encryption, and mitigates gradient-targeted attacks under a passive adversary threat model. It uses multi-dimensional packing to improve the batch parallelism efficiency of encrypted SIMD, and uses gradient clipping approximation to avoid gradient explosion. Experiments show that its model performance is close to non-secure schemes under homogeneous and heterogeneous data distributions, and scales linearly with the number of data holders and the number of time steps.