FitFEC: A Multi-Scale Transformer for Optimizing Packet Loss Recovery
Yining Li, Sirui Liu, Zongpeng Li, Liang Du, Ling Deng · 2025
We are witnessing more real-time network applications that demand high transmission reliability and low delay, for meeting desired Quality of Experience (QoE). Forward Error Correction (FEC) is widely employed to combat packet loss, but its effectiveness depends on accurate packet loss rate prediction. Existing forecasting models exhibit limitations in handling multiperiodicity and complex variations, particularly in terms of trend prediction consistency and forecasting conservativeness, leading to suboptimal performance of FEC strategies. To address these challenges, we propose FitFEC, a Multi-Scale Dynamic Adjustment Transformer that enhances packet loss prediction and optimizes FEC strategies. FitFEC integrates Adaptive Frequency Analysis for capturing periodic components, Trend-Detail Decomposition Transformer for improving trend accuracy, and Dynamic Prediction Adjustment to control prediction aggressiveness. Furthermore, we implement an FEC scheme based on QUIC to further enhance transmission efficiency. Extensive empirical studies demonstrate that FitFEC improves trend prediction accuracy, reduces retransmission rates, and reduces transmission latency, ultimately enhancing network performance and user experience.