关键词: 深度神经网络, 对抗攻击, 眩光效应, 模型鲁棒性, 对抗训练

Abstract:

With the continuous advancement of adversarial attack algorithms,the security risks that deep neural networks face are increasingly severe.Optical phenomena frequently occur in real-world scenarios,and the robustness against optical adversarial attacks directly reflects the safety of deep neural networks.Nevertheless,current research on optical adversarial attacks commonly encounters challenges such as optical perturbation distortion and optimization instability.To solve this problem,this paper proposes a novel optical attack method named AdvFlare,to help explore the effect of flare perturbations on the safety of deep neural networks.AdvFlare constructs a parameterized flare simulation model,which models the multiple attributes of the flare pattern,such as shape and color,with great simulation,on the basis of which this paper addresses the problems of adversarial perturbation distortion and convergence difficulties through strategies such as parameter space constraints,random initialization,and stepwise optimization.Experimental results indicate that AdvFlare can induce misclassification in deep neural networks with a significantly higher success rate compared to existing methods,while also offering a superior visual perturbation quality and stability.Furthermore,it is discovered that adversarial training using AdvFlare can markedly enhance the robustness of deep neural networks,in both the digital and physical world,providing valuable insights for improving model robustness in public transportation contexts.

Key words: deep neural networks, adversarial attacks, flare effect, model robustness, adversarial training