ET-SSM: Linear-Time Encrypted Traffic Classification Method Based On Structured State Space Model

Encrypted traffic classification involves categorizing data traffic that has been encrypted for privacy and security. In this domain, Transformers have been essential for successful pretraining. However, their quadratic time complexity makes them costly and slow for inference. While many works tried to reduce memory overhead and speed up inference, they often fail to match the classification performance of existing methods. To tackle this challenge, we present an Encrypted Traffic State Space Model (ET-SSM), utilizing recent advances in sequence routing based on state space models. The ET-SSM initially integrates SSM layers with a multiplicative gated architecture, which has proven effective in simplified sequence modeling frameworks. This structure learns static linear layers that do not account for pairwise traffic interactions. To overcome this limitation, we employ convolution layers to capture adjacent traffic context interactions while maintaining overall linear time complexity. Extensive experiments demonstrate that ET-SSM not only significantly reduces memory overhead and accelerates inference time but also achieves state-of-the-art classification results. This makes it a promising candidate for the pretraining model in encrypted traffic classification for realistic production environments. The source code will be released upon paper acceptance.