Method and system for improving adversarial defense capability of Bayesian neural network

Abstract

The invention provides a method and system for improving the adversarial defense capability of a Bayesian neural network, and relates to the technical field of Bayesian neural networks in the deep learning direction and adversarial defense, and the method comprises the following steps: a network training step: inputting initial data, and optimizing a cross entropy loss function and the KL divergence between parameter distribution and prior distribution to obtain a network model; training the initial data; and a robustness enhancement step: taking the parameters of the Bayesian neural network as input, calculating and optimizing the mathematical expectation of the spectral norm of the network parameter matrix, and reducing the sensitivity of the network to various noises according to the optimized mathematical expectation, thereby improving the adversarial defense capability of the Bayesian neural network model to the noises. According to the invention, the adversarial robustness of the Bayesian network model can be effectively improved, and the application prospect of the network model in practical problems is effectively enhanced; the output uncertainty of the network model is effectively reduced, the prediction of the network to the adversarial sample is more consistent, and the noise resistance of the network model is improved.

Description

technical field [0001] The invention relates to the technical field of Bayesian neural network and confrontation defense in the direction of deep learning, in particular, to a method and system for improving the Bayesian neural network confrontation defense capability. Background technique [0002] Bayesian neural networks are a special neural network architecture in the field of deep learning. Compared with the traditional neural network method, the parameters in the Bayesian neural network are random variables rather than constants. Therefore, even for the same input, the output of the network will vary across multiple runs, which is equivalent to sampling from a fixed probability distribution. This gives the Bayesian neural network the unique ability to model data uncertainty and model uncertainty, and therefore provides a certain degree of interpretability for the output of the network, making the Bayesian neural network in computer vision, natural language processing ...

AI Technical Summary

Problems solved by technology

However, since the ideal situation cannot be achieved in practical applications, there is still a lot of room for improvement in the robustness of Bayesian neural networks in practical problems.

Images