Model Entanglement for solving Privacy Preserving in Federated Learning
Keywords: Federated Learning, Privacy Preserving, Deep Learning, Data Representation.
Abstract: Federated learning (FL) is widely adopted as a secure and reliable distributed machine learning system for it allows participants to retain their training data locally, transmitting only model updates, such as gradients or parameters. However, the transmission process to the server can still lead to privacy leakage, as the updated information may be exploited to launch various privacy attacks. In this work, we present a key observation that the middle layer outputs, referred to as data representations, can exhibit independence in value distribution across different types of data. This enables us to capture the intrinsic relationship between data representations and private data, and inspires us to propose a Model Entanglement(ME) strategy aimed at enhancing privacy preserving by obfuscating the data representations of private models in a fine-grained manner, while improving the balance between privacy preservation and model accuracy. We compare our approach to the baseline FedAvg and two state-of-the-art defense methods. Our method demonstrates strong defense capabilities against mainstream privacy attacks, only reducing the global model accuracy by less than 0.7\% and training efficiency of 6.8\% respectively on the widely used dataset, excelling in both accuracy and privacy preserving.
Primary Area: alignment, fairness, safety, privacy, and societal considerations
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: 13534
Loading