Attention Mechanism-Based Adversarial Attack Against DRL

Deep Reinforcement Learning (DRL) seeks to optimize long-term future returns through learning policies based with deep learning models for achieving specific targets. However, current research has discovered that machine learning models are vulnerable to viciously crafted adversarial paradigms, which is the case for DRL that uses deep model learning policies. Gradient information is often used to produce adversarial perturbations to clearly observed states, thus disabling DRL. To develop a novel attack method to further test the flaws of DRL, we present a new adversarial attack based on the attention mechanism. Rather than using gradient information, we exploit the hidden features extracted in the DRL through attention manipulation to produce more effective adversarial examples. Both channel attention and pixel attention can be used to extract features to modify the clean state into an adversarial state. DQN is one of the state-of-the-art DRL models and is used as the target model to train the Flappybird gaming environment to assure continuous operation and high success rates. We performed comprehensive attack experiments on DQN and examined its attack performance in terms of reward, loss convergence, and computational complexity.