Khalid Adnan Alsayed

Hidden subgroup vulnerabilities in face analysis can lead to substantial performance disparities that traditional fairness metrics overlook.

Fairness assessments in ML are far more subjective than we thought: different metrics can completely flip your conclusion about whether a model is biased.

Aggregate accuracy can be dangerously misleading when evaluating facial recognition systems for law enforcement, obscuring significant disparities in error rates across demographic subgroups.