Accountable AI with ZKPs: Certifying Fairness and Explanations under model Confidentiality
summary
System builds a model commitment and produces per-query proofs of inference and explanation using zero-knowledge techniques, integrating the EZKL backend. LIME is reworked into four variants; the uniform+kernel variant yields ~1.1 minute proof generation, ~0.1 s verification and ~10 KB proofs for tabular credit models. Fairness certification splits offline (~1.5 minutes) and online (~4 s) proof steps with ~3 s verification. Proof cost rises quickly with hidden layers and parameters, showing feasibility for small models but scalability limits for large neural nets.
AI-written summary · about
chapters (15)
- 0:00Accountable AI with ZKPs
- 1:06AI risks and regulations
- 2:33Explainability limits
- 4:18Why explanations fail
- 7:15ZKPs for AI explanations
- 10:35Inference proof performance
- 13:12LIME local explanations
- 16:45ZKP explanation challenges
- 18:55ZKP-amenable LIME variants
- 21:05Model size vs ZKP overhead
- 22:34Fairness in confidential models
- 24:42Fairness proofs with ZKPs
- 26:07ZKP implementation challenges
- 26:51Future of accountable AI
- 27:47Q&A
related talks
33mZero knowledge machine learning with ezkl
22mPrivacy-preserving Computation in Plain Sight
16mBuilding Trust through Explainable AI
46mAI for High-Stakes Decision Making
26mPractical, Efficient ZKPs, Economic and Social Processes
41mSpotlight Talks
33mAn Introduction to Zero Knowledge Proofs
30mTowards a more transparent AI - Decrypting AI models using LIME