Evaluating Data Poisoning Vulnerability in Selective Reincarnation within c-MARL to Salt and Pepper Noise Attack
Harsha Putla, Chanakya Patibandla, Krishna Pratap Singh, Panduranga Naidu Nagabhushan · 2024
In the domain of cooperative multi-agent reinforcement learning (c-MARL), the deployment in safety-critical applications necessitates rigorous robustness testing. Despite the advancements in c-MARL algorithms facilitated by deep neural networks (DNN), the susceptibility of agent policies to adversarial examples poses significant risks. This study focuses on the susceptibility of selective reincarnation, a method leveraging historical agent experiences to expedite learning, in c-MARL systems against data poisoning attack targeting the offline teacher dataset, which consists of curated experiences collected during tabula rasa training. We tested the influence of salt and pepper noise on selective reincarnation and employed Hamming distance to quantify ranking disruptions. Our results reveal significant performance declines, with top-performing agent combinations (configurations) under clean conditions experiencing up to a $\mathbf{5 0 \%}$ decrease in performance under adversarial conditions. Furthermore, configurations with three or more agents exhibited a $\mathbf{1 0 0 \%}$ Hamming distance, signaling complete ranking reversals that critically affect reincarnation decisions. Our research not only enhances the understanding of the effects of data poisoning on c-MARL selective reincarnation but also contributes to the advancement of more secure multi-agent learning algorithms, ensuring robustness in critical applications.