UniGuardian: A Unified Defense for Detecting Prompt Injection, Backdoor Attacks and Adversarial Attacks in Large Language Models
UniGuardian proposes a single, training-free method called UniGuardian to detect prompt injection, backdoor attacks, and adversarial attacks by treating them as Prompt Trigger Attacks (PTA). It analyzes how removing specific words from a prompt affects the model's loss function during inference. This technique generates an uncertainty score to flag malicious prompts without needing any retraining.