Bufferless network traffic classification using reinforcement learning-based joint pruning-quantization

Deep learning (DL) has shown significant promise in network traffic classification, outperforming traditional methods such as port-based, payload-based, and earlier machine learning approaches. However, the complexity of DL models hinders their deployment on resource-constrained edge devices. Current lightweight traffic classification focus on global pruning and eliminating fully connected layers. Despite their effectiveness, these methods overlook two crucial aspects: the potential of layer-wise pruning for further complexity reduction; and the possibility of lowering computation precision. This paper proposes an approach that combines layer-wise pruning with quantization, utilizing reinforcement learning to efficiently explore the large design space for joint pruning-quantization. Our evaluations on the ISCX2016VPN, USTC-TFC2016, and ITC-NET-AUDIO-5 datasets demonstrate that the proposed method reduces computational complexity up to 8 times and parameter count by 3 times compared to existing methods, while maintaining acceptable accuracy levels. Overall, our proposed method classifies a single raw packet in 107 microseconds and requires only 24 kB of storage without buffering circuitry. This demonstrates its effectiveness in achieving high performance and efficiency for edge-based network traffic classification. The code is open-sourced at https://github.com/zafriazman/PruneQuant-TrafClassif .