Areas of interest include, but are not limited to, the following:
1. Privacy Leakage Detection
Violations of user privacy should be detected by monitoring and detection of information leakage. For example, poisoning analyses, where faulty user data is fed to the system and then detected emerging in a reconstructable fashion from the predictions, are an exemplar approach. We are interested in supporting tools that automate detection of privacy risks, particularly of black box systems.
AI systems should have safeguards and processes in place to actively prevent harms. Safeguards themselves need to be trusted and transparent. We are interested in novel approaches to linking monitoring to automated safety actions, as well as tools for the monitoring of safety measures themselves.
3. Fairness Issue Detection
Competing fairness objectives and frameworks can involve difficult tradeoffs or incompatibilities and there is no consensus measure of fairness. At the same time, observing fairness issues in an actionable, clear and timely fashion would help facilitate more collaborative discussions around appropriate remedies, motiviate speedy corrections, reduce potential aggregate harm, and provide greater accountability. We invite proposals that actively monitor for potential fairness issues, but also welcome work that specifies proposed fairness goals, measures and tradeoffs.
4. Interpretability and Explainability
The opacity of ML systems often adds weight to distrust in how systems are operating and can obscure potentially unrecognized harm. Automated monitoring that can uncover and describe the relative interpretability and defensibility of the learning patterns of ML systems can allow explainable systems to be recognized and promoted, and encourage opaque systems to be simplified or rebuilt to enhance trust. We welcome proposals for tools that provide interpretability and understandable explanations to ML decisions.
Systems that fail for users whose data is uncommon may create fairness risks for such individuals, and privacy leakage when those failures are observable. Analysis by fuzzing—whereby a continuously generated random stream of new files and edge cases are fed into a system to attempt to provoke errors—has proven to be one of the most powerful tools for building resilient operating systems, browsers and cloud environments, yet the same approach has made fewer inroads to large scale ML deployed systems. We are interested in proposals that can increase the stability of ML systems through continuous testing, such as fuzzing or other forms of automated analysis.
Reliable, accurate, replicable, auditable performance, constrained to a delimited purpose, are all hallmarks of a robust system. We are interested in tools that can monitor these dimensions, especially in cross-cutting ways that monitor the overarching integrity of performance.