Reduced Products of Abstract Domains for Fairness Certification of Neural Networks
summary
Libra is an open‑source static analyzer that certifies dependency fairness for ReLU feedforward classifiers by combining a cheap forward pre‑analysis over interval/symbolic/polyhedral domains with an exact polyhedra backward analysis on feasible partitions. A configurable budget splits when uncertain ReLU counts exceed a threshold; some partitions get excluded, others produce definite fair or biased input ranges. New reduced product domain construction shares bounds across domains, and experiments on a standard fairness dataset show improved precision while remaining scalable.
AI-written summary · about
chapters (10)
- 0:00Fairness certification intro
- 1:06Social impact and regulation
- 3:16Dependency fairness example
- 7:13Libra's architecture
- 10:24Backward analysis strategy
- 13:18Reduced products of domains
- 17:30Ongoing work framework
- 20:02Q&A: Gender perturbation
- 23:41Q&A: Performance specs
- 25:40Q&A: Variable impact
related talks
15mReduced Products of Abstract Domains for Fairness Certification of Neural Networks
15mPerfectly Parallel Fairness Certification of Neural Networks
56mAbstract Domains for Machine Learning Verification
22mFormal Security Analysis of Neural Networks using Symbolic Intervals
31mARENA: Enhancing Abstract Refinement for Neural Network Verification
14mGuarding the Privacy of Label-Only Access to Neural Network Classifiers via Formal(…)
14mStatic analysis of ReLU neural networks with tropical polyhedra
58mBuilding Trust and Safety in Artificial Intelligence with Abstract Interpretation