Privacy-Preserving Federated Recurrent Neural Networks

We present RHODE, a novel system that enables privacy-preserving training of and prediction on Recurrent Neural Networks (RNNs) in cross-silo federated learning setting by relying multiparty homomorphic encryption. RHODE preserves the confidentiality data, model, data; it mitigates attacks target gradients under passive-adversary threat model. propose packing scheme, multi-dimensional packing, for better utilization Single Instruction, Multiple Data (SIMD) operations With efficient processing, parallel, batch samples. To avoid exploding problem, provides several clipping approximations performing gradient experimentally show model performance with remains similar to non-secure solutions both homogeneous heterogeneous data distributions among holders. Our experimental evaluation shows scales linearly number holders timesteps, sub-linearly sub-quadratically features hidden units RNNs, respectively. best our knowledge, is first building blocks RNNs its variants, encryption setting.