ModelShield detects malicious ML models at runtime

Researchers have introduced ModelShield, a runtime behavioral detection system designed to identify malicious pre-trained machine learning models shared on public platforms. By combining application-level instrumentation of framework APIs with eBPF-based syscall tracing, the tool attributes observed behaviors to specific model operations and classifies them using MITRE ATT&CK techniques. Evaluations demonstrated high detection accuracy across multiple formats and frameworks, identifying numerous malicious payloads on Hugging Face that evaded existing static scanners. Open article →