Efficient Optimisation Framework for Convolutional Neural Networks with Secure Multiparty Computation

In recent years, deep learning has become an increasingly popular approach to modelling data due its ability detect abstract underlying patterns in data. Its practical applications have been limited, however, by privacy concerns, restricting use major sectors such as healthcare and banking. Secure multiparty computation (MPC) is a scheme which allows multiple parties perform joint computations over private data, while keeping the content of their secret. MPC can enable privacy-preserving machine learning, however current implementations are rarely applied practice prohibitively high cost performing thousands transmitting between parties. this paper we propose framework incorporating various optimisation approaches from wider field including batch normalisation polynomial approximation activation functions, evaluate performance when convolutional neural network (CNN), discussing trade-off each offers terms accuracy efficiency. We experiment with parametric (PPoly) activations deriving approximations functions allowing tune coefficients weights. will show that, shallow CNNs, application combination PPoly layer result faster convergence, testing exceeding that achieved unencrypted network, at longer running times.