Reduced Products of Abstract Domains for Fairness Certification of Neural Networks
summary
Tool analyzes ReLU feed‑forward classifiers and first runs a forward pre-analysis that splits the input space by counting uncertain ReLUs against configurable bounds L and U then hands feasible partitions to an exact backward analysis with polyhedra that propagates output regions back, projects away sensitive features and declares any nonempty intersection as bias. A reduced product domain combines Boxes, Symbolic, DeepPoly and Neurify, improving precision (≈11% in one config, ≈3% in others) and scales by parallelizing partitions across CPUs.
AI-written summary · about
chapters (10)
- 0:00Start
- 1:47Feed-Forward Neural Networks with ReLU Activations
- 2:55Dependency Fairness
- 3:15Recruiting Process
- 5:01Cheap Forward Pre-Analysis
- 8:37Proceed Backwards for each Feasible partitions for each possible Outcome
- 10:23Symbolic
- 11:10Precision-vs-Scalability
- 12:44Effect of Neural Network Structure
- 13:38Leveraging Multiple CPUs
related talks
28mReduced Products of Abstract Domains for Fairness Certification of Neural Networks
56mAbstract Domains for Machine Learning Verification
15mPerfectly Parallel Fairness Certification of Neural Networks
22mFormal Security Analysis of Neural Networks using Symbolic Intervals
28mStatic analysis of ReLU neural networks with tropical polyhedra
31mARENA: Enhancing Abstract Refinement for Neural Network Verification
14mGuarding the Privacy of Label-Only Access to Neural Network Classifiers via Formal(…)
31mAbstract Neural Networks