Vulnerability detection method and device for a deep learning system

Abstract

Embodiments of the present invention provide a method and device for detecting vulnerabilities in a deep learning system, wherein the method includes: selecting and activating several neurons from the deep learning system to be tested, guided by maximizing neuron coverage; The tensor expression of the element, and the tensor expression of the prediction difference of the deep learning system to be tested, construct an optimization function, and obtain several disturbances by maximizing the optimization function; if any disturbance in the several disturbances satisfies the preset condition, the detection sample is obtained based on the disturbance, and the vulnerability of the deep learning system to be tested is detected through the detection sample. The method and device provided by the embodiments of the present invention can effectively improve the neuron coverage rate, make the detection process more complete, and only need a deep learning system, so that the application scenarios of the vulnerability detection method are more extensive. In addition, a large number of samples can be generated, The detection efficiency of the vulnerability of the deep learning system to be tested is improved.

Description

technical field [0001] Embodiments of the present invention relate to the technical field of deep learning, and in particular, to a method and device for detecting vulnerabilities in a deep learning system. Background technique [0002] Currently, deep learning techniques are increasingly used in safety-related fields, such as autonomous driving and malware detection. The requirements for predictability and correctness in the above application fields have brought new challenges to the wide application of deep learning technology. Due to training data bias, model overfitting, etc., deep learning systems often exhibit unexpected or incorrect behavior in edge cases. In a security-critical environment, such incorrect behavior can lead to disastrous consequences. Therefore, the software security technology of the deep learning system is very important, which can effectively discover potential problems and defects in the deep learning system and ensure the reliability of the sys...

AI Technical Summary

Problems solved by technology

[0005] Embodiments of the present invention provide a method and device for detecting vulnerabilities in a deep learning system to solve problems existing in existing white-box testing methods

Images