A Robust and Resilient Quantized Federated Learning Framework over Data Imbalance and Lossy Communication Channels

Abstract

Federated Learning (FL) has emerged as a promising paradigm for collaborative machine learning across decentralized, resource-constrained environments. However, developing a resource-efficient FL framework that integrates compression techniques in non-independent and identically distributed (non-iid) data settings and lossy communication channels poses significant challenges, particularly for multi-agent robotic systems where heterogeneous devices—such as UnmannedAerial Vehicles (UAVs) and Unmanned Ground Vehicles (UGVs)—must operate under dynamic and contested network conditions. In this thesis, we present a robust and resilient quantized FL framework designed to overcome these challenges by simultaneously addressing network unreliability and model accuracy degradation under class imbalance. Our approach integrates an adaptive quantization and weight scaling strategy that optimizes resource efficiency while significantly boosting classification accuracy for minority classes. The proposed mechanism leverages Layer-wise Relevance Propagation (LRP) to assess class-relevant filter values at the client level, ensuring that critical information for minority classes is preserved during local training. Additionally, a server-side aggregation strategy is introduced to address the global class imbalance, resulting in a convergence rate of O 1?T under non-convex optimization settings. Robust communication is equally vital for efficient collaboration among distributed agents in unreliable networks. To address this, we designed a customized FL communication protocol supported by a high-fidelity simulation environment that co-simulates physics and network conditions. We developed a sliding window-based, velocity-aware synchronization middleware, SynchroSim, which minimizes the sim-to-real gap. Integrated as a QoS policy within the Robot Operating Systems (ROS) network, our protocol employs adaptive flow control and compressive sensing techniques. By dynamically adjusting the TCP receiver window size and selectively sampling vital information based on real-time network parameters (latency, RSSI, data rate) and agent velocity variations, a state-of-the-art deep neural network (TABNET) predicts optimal window sizes and compression ratios. We deployed our framework on an FL scenario comprising seven heterogeneous robotic agents and tested it over two different communication mediums. our proposed method enhances classification accuracy for minority classes by over 30% while reducing FL training latency by more than 56%. Moreover, our approach reduces packet loss by up to 15%, and improves retransmission rates by approximately 17%. The compressive sensing strategy achieves a payload reduction of nearly 56%.