AI for High-Stakes Decision Making
summary
Perturbation-based explainers like LIME and SHAP can be spoofed: an adversarial wrapper makes a model behave innocently on synthetic perturbations while still using race or gender on real cases, so explanations hide bias. A law‑student study shows the biased model is trusted by under 10% when its true features are visible but about 80% trust when the explanation omits race and shows correlates. Defenses include in‑distribution perturbations and robust/minimax explanation formulations.
AI-written summary · about
chapters (17)
- 0:00Intro
- 1:55Himas background
- 4:29Himas upcoming talk
- 6:29Blackbox vs interpretable models
- 9:57Limits of explained ability
- 13:37Effects of misleading explanations
- 19:15Future research
- 21:48Survey of explainability techniques
- 25:01How explainability techniques work
- 28:41Perturbation based methods
- 34:10The attacker
- 35:25Protections
- 36:11Two sides of a coin
- 38:01Other attacks
- 38:54Interesting connections
- 40:41More regularization is approach
- 43:49Technical challenges
related talks
35mWhy Do You Need an Explainable Ai?
15mIntroduction to Explainable AI
30mTowards a more transparent AI - Decrypting AI models using LIME
16mBuilding Trust through Explainable AI
28mManaging AI Risk: Beyond Explainability
37mExplaining model explainability
50mDehumanizing Agents: Why Explainability is Crucial in the LLM Era
01h 29mNavigating ML Explainability: From Theory to Implementation