Exploring Pillars of Trustworthy AI: Robustness and Fairness
summary
Shows how tiny, computed perturbations—derived from the loss gradient—break classifiers (using gradient-sign perturbation (FGSM)) and how adversarial training (a min‑max procedure, adversarial training min-max) raises robustness but cuts clean accuracy (example: standard model 95.2% → robust 87.3%, adversarial test 3.5% → robust 45.8%). Demonstrates LLM/multimodal jailbreaks via a garbage suffix-optimal jailbreak that forces unsafe completions, catalogs real dataset harms (Amazon hiring bias, speech‑to‑text WER gaps), and shows clear robustness ↔ fairness tradeoffs.
AI-written summary · about
chapters (9)
- 0:00Trustworthy AI pillars
- 1:26Global trust in AI
- 3:37Robustness in AI models
- 14:14Adversarial attacks defense
- 23:36Fairness and model bias
- 33:21Measuring fairness
- 36:38Robustness and fairness synergy
- 37:42Key takeaways
- 38:31Q&A
related talks
47mEthical Machine Learning
01h 24mAdversarial Examples
36mAdversarial Machine Learning: How Secure Is ML?
01h 05mNavigating the Security and Privacy Landscape of Modern AI
32mHow to Build Trustworthy AI with Open Source
49mA New Perspective on Adversarial Perturbations
47m1Provably Robust Deep Learning via Adversarially Trained Smoothed Classifiers
46mLessons Learned from Evaluating the Robustness of Defenses to Adversarial Examples