With the development and progress of artificial intelligence, machine learning has made major breakthroughs in many fields, such as face recognition, automatic driving and other fields have been widely used. However, while machine learning brings convenience to human beings, it also gradually exposes a series of security risks. Especially, the model represented by neural network is vulnerable to anti-sample attacks caused by difficult to detect disturbances due to its “black box” characteristics and unexplainability, thus affecting the robustness of the model. In order to meet this challenge, a variety of methods have been proposed to improve the robustness of models, among which adversarial training is considered to be an important means. Adversarial training regularizes model parameters by introducing adversarial samples, thus improving the robustness and generalization ability of the model. However, the existing adversarial training methods fail to effectively distinguish the intrinsic characteristics of adversarial samples in different scenarios, resulting in limited improvement of model robustness. To solve these problems, this paper proposes a false classification perceptual adversarial training method (EMRT) based on uncertainty measurement. Although the traditional Adversarial training method, which utilizes misclassification perception (MART), demonstrates effectiveness, it solely relies on prediction errors to discern erroneously classified samples, neglecting the impact of other categories, thereby posing challenges in attaining optimal performance. In this paper, we provide an improved scheme, that is, we first consider the effects of all classes of outputs of the countersample, and design an entropy based uncertainty measure (EMRT). This measure can effectively reflect the influence of misclassified and correctly classified samples. On this basis, a soft decision loss function is proposed, which further improves the robustness of adversarial training. We conducted experimental validation on the CIFAR-10 dataset, and the results show that the EMRT method significantly enhances the robustness of the model without sacrificing the prediction accuracy of natural samples.
Research on fault diagnosis model of robust control system based on machine learning
23.10.2024
563025 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
LPV Robust Filter Based Fault Diagnosis Method for Aeroengine Control System
Springer Verlag | 2024
|Railway turnout fault diagnosis method and system based on machine learning
Europäisches Patentamt | 2024
|