Subgroup-Robustness
Adversarial Robustness of Variational Autoencoders across Intersectional Subgroups.
This study examines the robustness of Variational Autoencoders (VAEs) against non-targeted adversarial attacks across demographic subgroups. Findings reveal that certain groups, like older women, are more vulnerable to misclassification due to adversarial perturbations.
Our study evaluates VAE robustness against non-targeted adversarial attacks across demographic subgroups (age and gender). It explores whether robustness disparities exist and what factors contribute, such as data scarcity and representation entanglement. The results show that certain subgroups, like older women, are more prone to misclassification due to adversarial perturbations. Analysis using downstream classifiers and latent embeddings highlights the underlying vulnerabilities in VAEs.