Local robustness verification method of neural network

Abstract

The invention relates to a local robustness verification method of a neural network, and the method comprises the steps: inputting the neural network obtained through training, and processing the semantics of the neural network to obtain a model; abstracting semantics of the neural network into operation on an abstract domain (or transformation on the abstract domain) of a mathematical structure based on the abstract interpretation; giving input and tiny disturbance; abstracting input and tiny disturbance into elements in an abstract domain, introducing a symbol propagation method, and converting a neural network local robustness verification problem into an abstract domain operation and judgment problem; and outputting an operation result so as to confirm whether the local robustness of the neural network is verified.

Description

technical field [0001] The invention relates to the technical field of software verification, in particular to a local robustness verification method of a neural network. Background technique [0002] Deep learning is a research direction in the field of machine learning. Deep learning has been developed in recent years and has made remarkable progress in speech and machine vision. Deep learning methods based on neural networks are the current development direction. Neural network is a kind of approximate artificial intelligence. It simulates the thinking activities of human neurons with hierarchically connected nodes, learns and trains input samples, and compares them with expected outputs. Through differential indicators such as loss functions, the network The connection weights are readjusted to improve the recognition accuracy of new samples. [0003] In the research field of computer vision, adversarial examples are a hot research direction. The so-called adversaria...

AI Technical Summary

Problems solved by technology

Generally speaking, for a specific example, given a sample, it is relatively simple to judge whether the sample is an adversarial sample, but it is difficult to formally verify whether the neural network has an adversarial sample under a certain condition, and the neural network itself The training and adjustment process is also very complex and time-consuming, so there is an urgent need to provide formal verification techniques to verify the local robustness of the neural network

However, due to the lack of interpretability of neural networks, formal verification technology has not yet seen a breakthrough in the local robustness verification of neural networks.

Images