.png){kind=icon}
Return
Automated Model Generation for Application Traffic Classification
.png){kind=icon}
PRE
AI
DOI：10.1016/j.knosys.2025.114740.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The continuous emergence of new applications presents significant challenges for network traffic classification tasks. Many studies have focused on constructing new deep neural network models to tackle this issue. However, designing and training these models not only require specialized expertise but also entail substantial time costs, making it difficult for these new methods to be implemented in real-world industrial settings. To address these challenges, we propose AutoMTC, a novel automated model generation framework based on neural architecture search. AutoMTC employs reinforcement learning to optimize the training of Recursive Neural Network (RNN) generators, effectively creating neural network architectures specifically designed for traffic classification. Experimental results reveal that compared to existing state-of-the-art algorithms, AutoMTC achieves Relative Error Rate Reductions (RERR) of up to 46% and 49% across three distinct datasets, highlighting the significant improvements and practical value of our approach. Additionally, by utilizing intelligent structure search, our method significantly reduces the labor costs associated with model design and tuning, facilitating the practical deployment of auto-generated, high-accuracy traffic classification models. AutoMTC is founded on the critical observation that CNN-based application traffic classification models comprise fundamental units, such as convolutional or pooling layers with varying parameters, and the optimal structure is attained through the selection and combination of these units. To effectively search for the most advantageous network architecture, we use a strategy based on reinforcement learning, which continuously generates model architectures and uses the classification performance to guide the search process with reward values.
Journal
K
IF:
7.6
Papers:
1.3W
Citations:
4.5W
Organization
Cited Papers
No cited papers available