Dynamic Quantization-Aware Neural Architecture Search for Real-Time Encrypted Traffic Classification in 5G Networks

This paper presents a dynamic quantization-aware neural architecture search (NAS) framework for real-time encrypted traffic classification in 5G networks. The framework jointly optimizes model architecture and quantization policies using reinforcement learning with hardware-in-the-loop feedback, addressing both accuracy and efficiency on edge devices. Unlike conventional static methods, the proposed system supports runtime bitwidth switching to adapt precision levels based on fluctuating traffic loads and hardware conditions. Extensive experiments were conducted on ISCX-VPN, USTC-TFC2016, and QUIC-5G datasets. The results show that the proposed approach achieves 94.2% accuracy on ISCX-VPN, 92.7% on USTC, and 89.4% on QUIC-5G, outperforming baseline methods while reducing inference latency by 28–42% and lowering energy consumption by up to 26%. The framework maintains robustness under low-precision constraints, with a mean Quantization Stability Score (QSS) of 0.91, compared to 0.82–0.87 for existing approaches. Hardware-specific optimizations provide additional gains, such as a 2.1 × speedup on Raspberry Pi 4, reduced latency on Jetson Xavier and lower energy consumption on Intel NCS2. The results of Ablation studies affirm that dynamic quantization, hardware feedback, and stability and mechanisms are an absolute necessity, improving accuracy by up to 3.1% and decreasing latency by 31% over sequential optimization. These results prove the efficiency of dynamic quantization-aware NAS to classify encrypted traffic and emphasize its prospects in broader 5G edge AI applications.