A Dual-Filtering Approach for Malicious Traffic Detection Against Label Noise and Class Imbalance Issues

作者：Xinkai Yi, Chenlu Zhuansun, Qinglin Yang, Pengdeng Li, Yuan Liu, Zhihong Tian · 发表于：International Conference on Data Science in Cyberspace · 年份：2025 · DOI：10.1109/dsc67331.2025.00026

The reliability of traditional model detection relies on accurately labeled training datasets in malicious traffic detection. However, the insufficient representativeness of minority class samples and the presence of mislabeled instances bring significant challenges to achieving reliable malicious traffic detection. To tackle these challenges, we propose a Dual-Filtering (Dual-F) framework that leverages neighbor-based label confidence and Gaussian kernel density. The framework incorporates a self-supervised learning mechanism, providing a principled and effective strategy for tackling the coexistence of label noise and class imbalance. In addition, we provide a detailed theoretical analysis to prove the effectiveness of the proposed approach in mathematical. Experimental results on the DoHBrw-2020 and CICIoT-2023 datasets show that our method achieves superior performance, with average accuracy (ACC) of 0.9174 and 0.9227, F1 scores of 0.9141 and 0.9207, and False Positive Rate (FPR) of 0.1125 and 0.0704, respectively. These results demonstrate that our method achieves average improvements of $61.75 \%$ in ACC, 141.9% in the F1 score, and an 85.05% reduction in FPR over baseline methods.