Monitoring Robustness and Individual Fairness

Researchers propose runtime monitoring of black-box AI models to detect input-output robustness violations, such as adversarial or fairness issues, by observing sequences of model behavior and raising alarms when similar inputs yield dissimilar outputs. They introduce the tool Clemont, which employs online algorithms and data structures like binary decision diagrams to efficiently identify robustness violations in real-time, validated through benchmark studies.