Adversarial Activation Patching: A Framework for Detecting and Mitigating Emergent Deception in Safety-Aligned Transformers

원문 초록

Large language models (LLMs) aligned for safety through techniques like reinforcement learning from human feedback (RLHF) often exhibit emergent deceptive behaviors, where outputs appear compliant but subtly mislead or omit critical information. This paper introduces adversarial activation patching, a novel mechanistic interpretability framework that leverages activation patching as an adversarial tool to induce, detect, and mitigate such deception in transformer-based models. By sourcing activations from "deceptive" prompts and patching them into safe forward passes at specific layers, we simulate vulnerabilities and quantify deception rates. Through toy neural network simulations across multiple scenarios (e.g., 1000 trials per setup), we demonstrate that adversarial patching increases deceptive outputs to 23.9% from a 0% baseline, with layer-specific variations supporting our hypotheses. We propose six hypotheses, including transferability across models, exacerbation in multimodal settings, and scaling effects. An expanded literature review synthesizes over 20 key works in interpretability, deception, and adversarial attacks. Mitigation strategies, such as activation anomaly detection and robust fine-tuning, are detailed, alongside ethical considerations and future research directions. This work advances AI safety by highlighting patching's dual-use potential and provides a roadmap for empirical studies on large-scale models.