SSGD: A Safe and Efficient Method of Gradient Descent

With the vigorous development of artificial intelligence technology, various engineering technology applications have been implemented one after another. The gradient descent method plays an important role in solving optimization problems, due to its simple structure, good stability, and easy implementation. However, multinode machine learning system, gradients usually need be shared, which will cause privacy leakage, because attackers can infer training data with information. In this paper, prevent leakage while keeping accuracy model, we propose super stochastic approach update parameters by concealing modulus length vectors converting it or them into a unit vector. Furthermore, analyze security demonstrate that our algorithm defend against attacks on gradient. Experiment results show is obviously superior prevalent approaches terms accuracy, robustness, adaptability large-scale batches. Interestingly, also resist model poisoning certain extent.