Attribute Leakage

Attribute leakage refers to the unintended disclosure of sensitive information (e.g., gender, race) from data or models, even when seemingly anonymized. Current research focuses on mitigating this leakage in various domains, including text, images, and point clouds, employing techniques like differential privacy, feature masking, and latent space manipulation within models such as StyleGAN and Graph Neural Networks. Addressing attribute leakage is crucial for ensuring privacy and fairness in machine learning applications, impacting areas like data security, algorithmic bias mitigation, and the responsible development of AI systems.