DQN-based Adaptive Network Architecture for High-Fidelity NGN Flow Classification

Next-generation networks (NGNs) are designed to handle the increasing demands of modern communication, driven by a surge in data traffic and a wide array of applications, including tactile internet, AR/VR, Video surveillance, etc. Efficient traffic flow classification is crucial in this dynamic network landscape for effective network management, analysis, and optimization, ensuring flow-specific security. Accurate and reliable classification allows for resource allocation, improving end-to-end Quality of Experience (QoE), and ensuring flow-specific security. Convolutional neural networks (CNNs) have been widely used for traffic classification. However, designing CNN models for diverse NGN flows presents new challenges aside from feature selection such as (i) hyperparameter optimization and (ii) model convergence for highly varying traffic patterns. To validate a CNN model, existing works use extensive testing for different hyperparameters, constrained by feature attribution values, initial state, search strategy, and action space of existing architectures. To address this, we develop a Neural Architecture Search (NAS) model based on Deep Q-Networks (DQN) with a quantized reward function to generate the CNN’s hyperparameter. The Q-learning agent adapts based on the rewards, leading to rapid convergence and high-fidelity traffic flow classification. Testing our implementation on the CMU-SynTraffic-2022 dataset reveals improved results that achieved a target accuracy of 96.4% within 30 episodes. This demonstrates the effectiveness and scalability of our proposed methodology compared to existing architectures for NGN traffic management and resilience.