Recent advances in Oblivious Neural Network Evaluation and Training

Adria Gascon

I will discuss recent advances in the problems of training and evaluating neural networks using Multi-Party Computation techniques, with an emphasis on variants of discrete neural networks considered so far. These include Binary Neural Networks, which are well known in the ML community, as well as custom variants specifically designed for MPC. I will also comment on efficient circuits for important aspects of training such as batch normalization and adaptive step size selection.