Combating Generative AI's Privacy Abuses
summary
Generative models already leak memorized training data, enable membership inference and prompt‑injection, and face poisoning risks; the panel cites a claim that 80% of enterprises adopt LLMs by 2026 and warns of deepfake incidents like the January 2024 Taylor Swift video. Practical mitigations include filtering PII, using differential privacy or synthetic data, red‑teaming and test suites, and isolating inference with confidential compute or future fully homomorphic encryption (current proofs take hours). Executive buy‑in and data hygiene are mandatory.
AI-written summary · about
chapters (9)
- 0:00Welcome and introduction
- 1:28AI growth and privacy concerns
- 10:07Privacy attacks in generative AI
- 22:32Mitigation strategies and regulations
- 33:11Defining responsible AI and ethics
- 36:43Incident response for AI breaches
- 39:12Audience Q&A on AI adoption
- 41:39Managing privacy in base models
- 43:03Call to action for responsible AI
related talks
13mNavigating Generative AI and Its Risks
45mOverview of Risks, Vulnerabilities and Paths to Safer Use of Generative AI
30mPrivate Generative Models: Balancing Privacy and Utility in AI
27mSafeguarding Generative AI Applications: Navigating Security Challenges and Best Pract...
46mGenerative AI: Cyber resiliency and risk
14mBuilding Transparent and Ethical Generative AI Models in High-Stakes Industries
18mAI Is Coming For Our Secrets
28mRound Table | Generative AI: Threat or Ally?