Securing Software Supply Chain: Problems, Solutions, and AI/ML Challenges
summary
Argues that malicious libraries matter more than CVEs because malware runs immediately, not just when exploited, and shows a proof‑of‑concept model that executes on load; recommends Sigstore/Rekor, cosign, SLSA and CycloneDX SBOMs with cdxgen multi‑signatures and VEX attestations for provenance and non‑exploitability claims. Describes a five‑dimension ML security framework (data, model, platform, compliance, human) being open‑sourced, demonstrates evasion attacks driven by epsilon perturbations, and warns that pre‑trained models and model hubs contain a small but rapidly growing set of malicious artifacts.
AI-written summary · about
chapters (19)
- 0:00Introduction
- 1:25Software supply chain attacks
- 3:14Key targets
- 6:04Attack surface mapping
- 9:58Frameworks for security
- 11:44Malicious vs vulnerable components
- 14:32Vetting third-party components
- 15:57SBOM transparency
- 18:27Securing build systems
- 20:56SBOM signing
- 22:19AI in cybersecurity
- 24:01Malwared ML attacks
- 26:29AI framework risks
- 27:52Evasion attacks
- 31:01AI model security
- 32:27Thales secure AI framework
- 35:12Attacker tactics
- 35:53Best practices
- 39:10Q&A
related talks
21mEnd-to-end secure ML Development
28mSecurity of Machine Learning
19mModel transparency for AI/ML security
28mFlawed ML Security: Mitigating Security Vulnerabilities in Data & Machine...
19mWe Know Security but How Do We Secure GenAI End-to-End?
43mLogically Securing the Illogically Logical Use of Large Language Models
27mA Gentle Introduction to MLSecOps
31mSecuring the Future: Navigating AI Vulnerabilities and Evolving Security Practices