Reproduce “Adversarial Learning of Privacy-Preserving and Task-Oriented Representations” on OpenResearch

About this paper

Taihong Xiao, Yi-Hsuan Tsai, Kihyuk Sohn, Manmohan Chandraker, Ming-Hsuan Yang

An adversarial learning framework creates deep representations that balance utility for a target task with resistance to model inversion attacks. This approach effectively prevents image reconstruction from features while maintaining task performance, offering tunable control over the utility-privacy trade-off without requiring explicit private attribute labels.

Key insights

• Adversarial training against reconstruction loss effectively discards private information from deep features without requiring explicit labels for private attributes.
• The framework provides a direct mechanism to control the utility-privacy trade-off, allowing for adaptable deployment based on specific application requirements.
• The methodology aligns with the Information Bottleneck principle, implicitly optimizing to compress task-relevant information while maximizing uncertainty regarding private details.

Summary by alphaXiv.