Research on Adversarial Defense Mechanisms for Deep Learning-based Vision Models

Deep learning-based vision models have achieved remarkable success but remain vulnerable to adversarial attacks, which compromise their security in real-world applications. This paper proposes an adversarial defense mechanism combining a feature attribution-based attack generation method and a joint defense strategy using image denoising and feature correction. The proposed method enhances attack transferability by suppressing critical neuron activations and improves model robustness through noise simulation and feature adjustment. Experimental results show that the approach achieves superior defense success rates across multiple attack scenarios while maintaining model accuracy. Finally, the challenges and future directions of adversarial defense are discussed.