Layer-wise Federated Learning for Mobile Networks

Data privacy concerns have become a significant challenge in modern communication networks, where network clients generate diverse types of potentially sensitive information. Federated learning (FL) emerges as a privacy-preserving solution that allows distributed model training without centralizing sensitive data, thus enabling collaborative learning across decentralized networks. To collaboratively learn a global model in FL, participating users and a central entity periodically exchange information. Real-world network conditions, particularly varying channel quality, present obstacles to effective information exchange in federated learning implementations. In this paper, we introduce a federated learning framework that takes into account the channel quality variations among users and enhances resource utilization with reduced compromise to model predictive performance. Our proposed framework capitalizes on the observation that neural network layers exhibit diverse levels of sensitivity to noise. Our approach leverages this observation by selectively transmitting per-layer model updates based on channel conditions. This strategy not only protects critical model components from adverse channel effects but also conserves transmission resources. We demonstrate that our method achieves comparable accuracy to traditional approaches while significantly reducing communication overhead. To showcase its potential, we validate our method using two datasets. Our simulation results demonstrate promising performance in terms of both resource utilization and accuracy.