Defending Against Adaptive Adversarial Attacks: A Reinforcement Learning-Based Approach to Deep Neural Network Robustness

Adversarial attacks generated through reinforcement learning present a significant challenge to the robustness of deep neural networks due to their adaptive and non-deterministic nature. Traditional defense strategies often rely on the consistency of known attacks, which limits their effectiveness against more dynamic threats. In this work, we propose the Adaptive Reinforcement-Based Defense (ARBD), a novel defense framework designed to mitigate the impact of reinforcement learning-based adversarial attacks. ARBD integrates adversarial training with a pre-processing technique known as feature squeezing to reduce the influence of small perturbations by applying Bit Depth Reduction and Spatial Smoothing (Median Filtering). This combination enables the model to learn robust representations while minimizing the effect of subtle input manipulations. Experimental results show that ARBD reduces the RLAB attack success rate to 0.31% on MNIST and as low as 2% on CIFAR-10. Furthermore, it maintains or improves clean accuracy across all tested models, achieving up to 99% on MNIST and 98% on CIFAR-10.