Roman Belaire

Roman Belaire, Pradeep Varakantham, Thanh Nguyen et al.

Deep Reinforcement Learning (DRL) policies have been shown to be vulnerable to small adversarial noise in observations. Such adversarial noise can have disastrous consequences in safety-critical environments. For instance, a self-driving car receiving adversarially perturbed sensory observations about nearby signs (e.g., a stop sign physically altered to be perceived as a speed limit sign) or objects (e.g., cars altered to be recognized as trees) can be fatal. Existing approaches for making RL algorithms robust to an observation-perturbing adversary have focused on reactive approaches that iteratively improve against adversarial examples generated at each iteration. While such approaches have been shown to provide improvements over regular RL methods, they are reactive and can fare significantly worse if certain categories of adversarial examples are not generated during training. To that end, we pursue a more proactive approach that relies on directly optimizing a well-studied robustness measure, regret instead of expected value. We provide a principled approach that minimizes maximum regret over a "neighborhood" of observations to the received "observation". Our regret criterion can be used to modify existing value- and policy-based Deep RL methods. We demonstrate that our approaches provide a significant improvement in performance across a wide variety of benchmarks against leading approaches for robust Deep RL.

Roman Belaire, Arunesh Sinha, Pradeep Varakantham

Deep Reinforcement Learning (DRL) policies are highly susceptible to adversarial noise in observations, which poses significant risks in safety-critical scenarios. The challenge inherent to adversarial perturbations is that by altering the information observed by the agent, the state becomes only partially observable. Existing approaches address this by either enforcing consistent actions across nearby states or maximizing the worst-case value within adversarially perturbed observations. However, the former suffers from performance degradation when attacks succeed, while the latter tends to be overly conservative, leading to suboptimal performance in benign settings. We hypothesize that these limitations stem from their failing to account for partial observability directly. To this end, we introduce a novel objective called Adversarial Counterfactual Error (ACoE), defined on the beliefs about the true state and balancing value optimization with robustness. To make ACoE scalable in model-free settings, we propose the theoretically-grounded surrogate objective Cumulative-ACoE (C-ACoE). Our empirical evaluations on standard benchmarks (MuJoCo, Atari, and Highway) demonstrate that our method significantly outperforms current state-of-the-art approaches for addressing adversarial RL challenges, offering a promising direction for improving robustness in DRL under adversarial conditions. Our code is available at https://github.com/romanbelaire/acoe-robust-rl.

Roman Belaire, Arunesh Sinha, Pradeep Varakantham

When a user asks a language model something harmful, is it a genuine attack or a misunderstood but well-meaning question? This ambiguity is one of the central challenges of LLM safety. A model that assumes the worst harms legitimate users; one that assumes the best is easily exploited. The problem is compounded in multi-turn dialogue, where an attacker's true intent may only reveal itself gradually across many exchanges, yet existing safety frameworks apply a contextual bandit treatment, ignoring the trajectory of the conversation. To that end, we propose Dialogue Critic Guided Sampling (DCGS), a framework that addresses this by inferring user intent at every turn of dialogue. Instead of applying a fixed rule about what is or is not safe, DCGS learns what the user's intent is likely to be based on the full conversational history and generates responses accordingly. Formally, we model adversarial dialogue as a Markov Decision Process and learn value and regret-based critics at both the individual token and utterance (full response) levels, scoring candidate responses via an action-value critic. We prove that this inference-time reweighting approximates exponential tilting of the base policy, guaranteeing improvement in expected return for any finite candidate pool, a property that group-relative objectives do not exhibit. Evaluated on CARES-18k, WildJailbreak, Redbench, and Harmbench, DCGS outperforms strong robust baselines and frontier models on adversarial dialogue tasks. DCGS also transfers to frontier models, improving their robustness without fine-tuning.