Deep learning system robustness enhancement method and device based on particle swarm optimization

Abstract

The invention discloses a deep learning system robustness enhancement method and device based on a particle swarm algorithm. The method comprises the following steps: acquiring an image data set and sorting and classifying the image data set into a training set and a test set; performing adversarial attack pre-training on a clean deep learning model by using the training set; constructing a robust evaluation index to evaluate the trained deep learning model; constructing zero-order and first-order loss function measurement for evaluating the relationship between the test set and the trained deep learning model; taking the first-order robustness measurement index as an optimization objective function, collecting a part of benign images from the test set as initial particles, initializing the positions and speeds of the particles, taking the optimization objective function as a maximum target, and carrying out iterative updating on the positions and speeds of the initial particles by adopting a particle swarm optimization algorithm to generate a test sample; retraining the deep learning model by using the generated test sample and the test data set to obtain a robust defense model; and evaluating the attack sensitivity of the model after defense.

Description

technical field [0001] The invention relates to the field of deep learning security, in particular to a method and device for enhancing the robustness of a deep learning system based on a particle swarm algorithm. Background technique [0002] With the rapid development of deep learning, deep learning is the core driving force behind the unprecedented breakthroughs in many real-world problems. [0003] Although deep learning has made great progress, deep learning systems are vulnerable to adversarial attacks. The adversarial examples of adversarial attacks are slightly perturbed on the basis of the original samples, which are slightly different from the normal input of human observers. , but it is easy to trick a deep learning system into making wrong decisions. The existence of adversarial attacks not only damages the reliability and robustness of deep learning systems, but also leads to security problems in countless deep learning systems in applications such as face reco...

AI Technical Summary

Problems solved by technology

However, formal verification techniques are usually expensive and only applicable to limited model structures and properties

[0006] System testing aims to generate test samples that can expose the vulnerabilities of deep learning models, and then, the test samples can improve the robustness of the model by retraining the model; The test sample is not directly related to the robustness of the model, which will limit the robustness improvement of the retrained deep learning model

Images