Enabling locality-sensitive machine learning towards low predictive overhead in flow classification

Abstract

The expansion of computer networks and the proliferation of their applications have heightened the necessity for accurate network traffic identification to enhance resource allocation and mitigate congestion, emphasizing flow classification’s pivotal role. Traditional methods, notably deep packet inspection, have encountered challenges due to the rise in encrypted traffic, diminishing their efficacy. In response, machine learning approaches, including decision trees, have been adopted for their ability to classify flows without necessitating access to the payload, thereby enhancing classification efficiency. Nonetheless, these machine learning-based models introduce considerable overhead in category determination. To address this issue, this paper introduces a locality-sensitive flow classification (LSFC) method, leveraging locality-sensitive hashing to efficiently partition and group traffic, thereby expediting the classification process and minimizing overhead. LSFC adeptly balances classification accuracy and processing delay, catering to the specific demands of various scenarios. Comprehensive evaluations demonstrate that LSFC significantly outperforms traditional machine learning models in inference speed by over 50% and improves upon deep packet inspection techniques in handling encrypted traffic, achieving a reduction in classification delay exceeding 15% without compromising classification accuracy.