Abstract
With the rapid advancement of information technology, the Internet, as the core infrastructure for global information exchange, faces increasingly severe security challenges. However, traditional network traffic detection methods typically focus solely on the local features of traffic, failing to comprehensively consider the global relationships between traffic flows. This limitation results in poor detection performance against multi-flow coordinated attacks. Additionally, the inherent imbalance in real-world network traffic data significantly hampers the performance of most models in practical scenarios. To address these issues, this paper proposes a network traffic detection method based on data interpolation and contrastive learning (TICL). The method employs data interpolation techniques to generate negative samples, effectively mitigating the data imbalance problem in real-world scenarios. Furthermore, to enhance the model's generalization capability, contrastive learning is introduced to capture the differences between positive and negative samples, thereby improving detection performance. Experimental results on two publicly available real-world datasets demonstrate that TICL significantly outperforms existing intrusion detection methods in large-scale data scenarios, showcasing its strong potential for practical applications.