Insane in the Supply Chain: Threat modeling for...
summary
Examines concrete attack paths: web-scale data poisoning that flips classifiers with 0.1% poisoned for $60, RAG poisoning achieving 90% success from five injected texts, and a compromised Hugging Face conversion service that exfiltrated a converter token and pushed malicious pull requests, plus practical exploits in pickle, HDF5/Keras Lambda layers, TensorFlow features, ONNX directory traversal and package-squatting (torchtriton). Shows deployed demos and live scans, recommends checksums, sandbox execution, continuous monitoring and model signing/lineage as fixes.
AI-written summary · about
chapters (8)
- 0:00Introduction and scope
- 3:55Supply chain security
- 7:10Data poisoning risks
- 13:10Model hijacking and exploits
- 25:56Securing model downloads
- 29:53Tooling and dependency issues
- 38:10Advanced attack vectors
- 41:48Mitigation and conclusions
related talks
50mMalice In Chains: Supply Chain Attacks Using Machine Learning Models
45mThe Dark Side of AI: The Hidden Supply Chain Risks in Open...
49mWhat Hacking the Planet Taught Us About Defending Supply Chain Attacks
33mUnderstanding the Threats and Attacks on Data Science Applications and Models
56mHidden Risks in Open-Source Code and AI Models
37mSponsored Session: The Simple, Yet Lethal, Anatomy of a Software Supply Chain Attack
34mThe Simple, Yet Lethal, Anatomy of a Software Supply Chain Attack
42mTracking Attackers in Open Source Supply Chain Attacks: The New Frontier