ResearchBenchmarksBackdoor Detection on ChatDoctorFollow0.0058F1-scoreLLMScan-0.000230.0013230.0028760.004428Dec 22, 2025Evaluation ResultsMethodMethodLinksF1-scoreClean Sample AccuracyPoison Sample AccuracyClean Weight AccuracyPoison Weight AccuracyLLMScanDefense Paradigm=Infer...Defense Paradigm=Inference-time behavior analysis2025.120.0058——98.5625.48ONIONDefense Paradigm=Data-...Defense Paradigm=Data-level filtering2025.12010——PEFTGuardDefense Paradigm=Weigh...Defense Paradigm=Weight-level inspection2025.120——1000