Lessons Learned from Developing Secure AI Workflows
summary
Google describes SAIF as a practical checklist: concrete risks include data poisoning via user feedback (Gmail spikes from spammers), model backdoors hidden in serialized Python layers (Hugging Face incidents), and model exfiltration mitigated by bucketed outputs for vision classifiers. Defenses used or recommended are dataset sanitization, access control, model signing for integrity, differential privacy for PII, red‑team prompt‑injection tests, expanding bug bounties, and tooling like QLoRA/LoRA for fine‑tuning; five priorities are filter the input, filter the output, enforce least privilege.
AI-written summary · about
chapters (12)
- 0:00Secure AI workflows intro
- 2:14SAIF framework overview
- 3:42AI system mental model
- 7:16Data risks and mitigation
- 10:07Infrastructure risks for AI
- 14:07Model risks and backdoors
- 18:46Input/output security challenges
- 26:10Secure AI integration
- 30:51AI governance and compliance
- 35:26Top five security tips
- 37:54Q&A: Model filtering & explainability
- 41:25Q&A: Data poisoning & small enterprises
related talks
31mSecurity Considerations for Services Using AI Models
57mAI Applications Security Puzzle
09mSecuring Generative AI Workloads
12mSecuring AI-Driven Finance
34mInvisible Infiltration of AI Supply Chains: Protective Measures from Adversari...
39mSecuring AI at Scale: Practical Defenses against Prompt...
55mAttacking And Protecting Artificial Intelligence
21mNavigating the Intersection: AI’s Role in Shaping the Secure Open Source Software Ecos...