From Attacker's Eyes: Preventing Attribute Privacy Leakage in User Behavior Data

Abstract Attribute inference attacks exploit the correlation between user behavioral data and private information to infer private attributes from behavioral data, leading to privacy leakage risks. Typically, defense against such malicious privacy inference is perceived as conflicting with attacks. However, this paper revisits representative attribute-privacy-preserving methods within an attack framework, revealing that their optimization objectives align with adversarial attacks with only minor differences. This unified attack lens allows us to assign precise meanings to their protection mechanisms and facilitates comparisons of their strengths and limitations. Based on our observations, we outline the major issues in applying attack to privacy preservation, such as selecting perturbed and target data. We then identify an appropriate attack optimization objective and present a poisoning-based privacy-preserving method to address these issues. Finally, we prove the privacy-preserving performance through theoretical analysis and experimental evaluation, and explore its enhancements over existing methods as well as the challenge of attacker decryption under open principle.