Rank-Based Traffic Classification for QoS: In-Network Inference Approach

Recently, the increase of low-latency applications has led to the need for faster data transmission. Meanwhile, in-network inference, which utilizes neural networks (NNs) in the data plane for traffic analysis (e.g., traffic classification, in-trusion detection), is gaining attention for its line-rate processing capability. This approach offers a solution for low-latency tasks by enabling simultaneous traffic analysis and packet processing within the network. However, existing approaches primarily focus on traffic analysis and overhead reduction without considering flow priorities, which can not satisfy sufficient performance for diverse applications. To address this, we propose rank-based traffic classification that considers both classification accuracy and quality of service (QoS) of each flow. First, we suggest a utility function that represents both task accuracy and the degree of satisfaction about flow requirements. Next, we use early-exit inference to meet flow requirements by deciding whether to stop inference procedure or continue to the subsequent steps based on the utility function. Experimental results show that our approach reduces flow completion time compared with baseline, while achieving comparable classification accuracy.