Communication-Efficient Federated Learning via Gradient Distillation
Keywords: Federated Learning, Edge Intelligence
Abstract: Federated learning revolutionizes collaborative model training across decentralized edge devices, ensuring privacy by avoiding direct data sharing. However, the frequent exchange of model updates introduces a significant communication overhead. The conventional FL process involves transmitting the differences in parameters between old and new models, resulting in redundant gradient communications due to the intricate interplay between model parameters and network architecture. Even minor adjustments to parameters necessitate the retransmission of entire models. In this paper, we introduce a groundbreaking concept known as \textit{gradient distillation}, which decouples model parameters from network architecture, enabling the transmission of only essential information needed for synchronization. By leveraging gradient distillation, we approximate gradient disparities into a synthetic tensor sequence, allowing the recipient to reconstruct the sender's intended model update. This approach eliminates the need to transmit the entire set of raw parameter differences, offering a highly promising solution for achieving greater communication efficiency while without significant accuracy degradation. Experimental results demonstrate that our approach achieves an unprecedented level of gradient compression performance, surpassing widely recognized baselines by an impressive margin of orders of magnitude.
Primary Area: societal considerations including fairness, safety, privacy
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/2024/AuthorGuide.
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: 200
Loading