Year: 2025
Type: conference-paper
Abstract: Federated learning is a new machine learning approach that focuses on privacy protection and allows models to be trained collaboratively using distributed data. The key idea is to update the global model through local computations in each iteration, thus avoiding direct transfer of raw data. However, when model updates with lots of parameters need constant communication, problems start to appear. Its overhead increases significantly, putting a heavy burden on the client system. Device heterogeneity exacerbates these obstacles. To tackle these limitations, we propose CDFedKD, a federated learning framework that merges adaptive mutual knowledge extraction with dynamic gradient compression. Although these techniques share some similarities with previous methods, their use in CDFedKD brings about new advantages. Mutual knowledge distillation enhances gradient compression efficiency, makes model updating easier, and cuts down on communication needs. The adaptive mechanisms within the framework boost robustness against device variability in real time. Empirical results show that combining the framework with the baseline method reduces model loss significantly and doesn't add much to training time. Experiments on various datasets and devices show that CDFedKD can effectively reduce communication inefficiencies and handle device heterogeneity. This makes a broader range of federated learning applications possible.
Authors: Qilong Feng, Li Luo
Institution: Chongqing University of Posts and Telecommunications
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4412789879