Privacy Preserving Generative Feature Transformation
Keywords: Data-Centric AI, Privacy-Preserving, Generative Model
Abstract: Data-Centric AI (DCAI) aims to use AI to get better data for better AI. Feature transformation, as one of the essential tasks of DCAI, can augment the data representation and has garnered significant attention. Existing methods have demonstrated state-of-the-art performance on advancing predictive tasks. However, these methods can lead to serious privacy leakage. For example, sensitive features in original data can be inferred by models trained on transformed data, exposing vulnerabilities in the privacy-preserving capabilities of these methods. To address this issue, we introduce a privacy-preserving feature transformation framework that transforms data representation while preserving privacy from a generative modeling perspective. Specifically, our framework includes two phases: 1) privacy-aware knowledge acquisition and 2) privacy-preserving feature space generation. In the knowledge acquisition phase, we develop an information bottlenecks guided reinforcement learning system to explore and collect privacy-aware feature sets as a knowledge base in token sequence form. In the feature space generation phase, we develop a generative model to encode the knowledge base into a privacy-aware latent space, where the best latent representation is identified and decoded into the optimal privacy-preserving feature space. We solve the optimization via projected gradient ascent that maximizes predictive performance and minimizes privacy exposure. Finally, we present extensive experiments on eight real-world datasets to evaluate how our method can navigate both performance and privacy. The code is available at https://anonymous.4open.science/r/anonymous-2B53/.
Primary Area: other topics in machine learning (i.e., none of the above)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 11711
Loading