Robustness of Neuro-Symbolic Reinforcement Learning to Adversarial Scenarios

Neuro-symbolic reinforcement learning approaches represent a novel paradigm in artificial intelligence that shows great potential in solving complex issues that traditional deep-learning techniques can only partially address. Despite their relevance and popularity, little to no attention has been directed to the study of their overall security and robustness to adversarial scenarios, due to their recent introduction in the literature. Furthermore, the evergrowing rate of adoption of AI-based techniques in real-world applications strongly emphasises the urgency for a careful assessment of their security implications aimed at reducing the risk of ethical, social and economic damages potentially arising from their use. Building on these observations, my research aims to evaluate the robustness of state-of-the-art neuro-symbolic reinforcement learning algorithms to both known and novel attacks capable of posing a significant threat to real-world deployments and investigating the potential mitigations that could be used to prevent such attacks.