FL-DA-ARQ: A FEDERATED LEARNING FRAMEWORK FOR DELAY-AWARE AUTOMATIC REPEAT REQUEST IN DISTRIBUTED PACKET TRANSMISSION AND TIMEOUT PERFORMANCE EVALUATION

Abstract

 Automatic Repeat Request (ARQ) protocols play a very important role in the reliable transmission of packets, but have the disadvantage of causing increased latency and retransmission overheads in both heterogeneous and distributed networks. Current delay-aware ARQ (DA-ARQ) models optimize strategies of timeouts and retransmission on a local basis, without coordination and scaling across a multi-node system. In order to fill this gap, we introduce FL-DA-ARQ, the first federated learning system that combines delay-conscious ARQ and distributed intelligence. As opposed to the traditional centralized or independent ARQ schemes, FL-DA-ARQ uses BiLSTM and CNN-LSTM predictors that predict the probability of timeouts and dynamically adjust retransmission windows. An improved version of a federated averaging (FedAvg) scheme with differential privacy (DP) and gradient compression allows model training on a federated scale, maintains overall data privacy, and reduces the operational overhead on communication. Experimental measurements prove that FL-DA-ARQ establishes a 97.3% packet delivery ratio, 45.2-ms mean latency, and 1.2 retransmissions of a packet, a 23% improvement in packet delivery ratio, and an 18 percent decrease in transmission delay in comparison with standard DA-ARQ. The framework is also highly scalable and robust to non-IID data distributions, and is thus applicable to large-scale privacy-preserving and delay-sensitive distributed network applications.