PsychoPass: Geometric Profiling of Multi-Turn Adversarial LLM Conversations (arxiv.org)

Researchers have introduced PsychoPass, a novel framework designed to address vulnerabilities in large language models (LLMs) during multi-turn conversations. Current security methods focus on individual conversation turns, but adversarial attacks often span multiple exchanges, resulting in a mismatch in detection capabilities. PsychoPass rethinks this approach by modeling conversations as paths in a geometric representation space, enabling the prediction of malicious intent based on early geometric features within conversation trajectories. The system demonstrates a high accuracy rate in classifying potential attacks, even when controlling for length-based confounders. The significance of PsychoPass lies in its ability to identify adversarial patterns early in a conversation, enhancing the robustness of LLM safety measures. This proactive monitoring approach can potentially mitigate risks before harmful content is generated, marking a substantial advancement in LLM security. The findings indicate that adversarial intents may leave discernible geometric fingerprints that are consistent across various encoding methods, providing a pathway for real-time protective mechanisms that could be integrated into existing AI systems.