ResearchBenchmarksBackdoor Detection on PII-MaskerFollow0.8926F1LLMScan-0.0357040.2052980.44630.687302Dec 22, 2025Evaluation ResultsMethodMethodLinksF1Clean Sample AccPoison Sample AccClean Weight AccPoison Weight AccLLMScanDefense Paradigm=Infer...Defense Paradigm=Inference-time behavior analysis2025.120.8926——85.9889.2ONIONDefense Paradigm=Data-...Defense Paradigm=Data-level filtering2025.120.06561003.39——PEFTGuardDefense Paradigm=Weigh...Defense Paradigm=Weight-level inspection2025.120——1000