EPIC: Ensembled Poisoning Identification and Classification for LLM Backdoor Defense

In the realm of artificial intelligence, Large Language Models (LLMs) stand as titans of natural language understanding and generation, driving innovations across diverse applications. However, their ascendancy brings forth challenges, notably vulnerability to data poisoning attacks; malicious alterations designed to compromise model integrity. This work introduces a novel approach to safeguard LLMs against these malicious threats. This paper delineates a two-pronged methodology: initially deploying an ensemble of traditional machine learning models to detect discrepancies in predictions between clean and poisoned models, followed by an enhanced technique that trains models to discern the “poison status” directly from data inputs. Through meticulous experimentation, this work not only unveils the susceptibility of LLMs to poisoning but also demonstrates the efficacy of ensemble methods and poison detection in fortifying model defenses. Our research aims to improve security and robustness of AI systems.