ResearchBenchmarksBackdoor Detection on SafetyLLMFollow0.8083F1 ScoreLLMScan-0.0323320.1859090.404150.622391Dec 22, 2025Evaluation ResultsMethodMethodLinksF1 ScoreAccuracy (Clean Sample)Accuracy (Poison Sample)Accuracy (Clean Weight)Accuracy (Poison Weight)LLMScanDefense Paradigm=Infer...Defense Paradigm=Inference-time behavior analysis2025.120.8083——93.1371.88ONIONDefense Paradigm=Data-...Defense Paradigm=Data-level filtering2025.120.100810.0531——PEFTGuardDefense Paradigm=Weigh...Defense Paradigm=Weight-level inspection2025.120——1000