Fast Learning Enabled by In-Network Drift Detection

Bruno Missi Xavier, Magnos Martinello, Celio Trois, Brenno de Mello Alencar, Ricardo Araújo Rios

Abstract: The widespread adoption of Machine Learning (ML) is leading to an increase in processing demands. Dealing with the growing volume of data poses a significant challenge in providing accurate classification services using ML models. Offloading ML tasks to network switches presents an opportunity to tackle this challenge, offering high throughput and low latency. Nonetheless, network devices encounter limitations in resources, and programmable languages like P4 lack support for basic operations, necessary for the ML methods, including floating-point arithmetic and native repetition structures. In this paper, we investigate the use of drift detection ML models to enhance the accuracy of in-network traffic classification. The novelty lies in designing drift detection based on bitwise operations, which are well-suited for implementation within the data plane. As a proof-of-concept, we implement drift detection using the P4 language on BMv2 switches, validated with a dataset of over 2 million samples. Our results demonstrate a significant increase in classification accuracy with drift detection, while maintaining line-speed operation and quickly adapting to changes in traffic patterns.