Perfectly Parallel Fairness Certification of Neural Networks
summary
A perfectly parallel static analysis certifies dependency fairness for feed‑forward ReLU networks by combining a cheap forward pre-analysis (interval abstract domain) that recursively partitions inputs with a bounded-path backward analysis that fixes at most an upper bound of 2 non-fixed ReLUs in examples. The tool flags real bias in a credit‑screening model where age alone flips decisions, detects injected bias in experiments, and shows precision/speed tradeoffs where larger queried input spaces hurt scalability most.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:33Neural Networks
- 2:10Fairness
- 3:23Static Analysis
- 4:13Backward Analysis
- 5:37Exponential Explosion
- 6:11Solution
- 8:15Example
- 12:03Experiments
- 13:36Scalability
- 14:45Conclusion
related talks
28mReduced Products of Abstract Domains for Fairness Certification of Neural Networks
15mReduced Products of Abstract Domains for Fairness Certification of Neural Networks
22mFormal Security Analysis of Neural Networks using Symbolic Intervals
59mDifferential Verification of Deep Neural Networks
23mRemoving Unfair Bias in the Machine Learning Pipeline
42mFairness in Machine Learning: From Theory to Practice
35mOpen Source Fairness
34mInterpretability and Algorithmic Fairness