关键词: 噪声标签监测, 噪声权重, 加密恶意流量, DoH流量, 差分训练

Abstract: Processing noisy datasets remains a challenge for training and evaluating data driven encrypted malicious traffic detection models. A noise label detection method based on KRPDDT was proposed, which used differential training to train two identical models simultaneously, extracted the training losses of samples in the two models, and detected noise samples based on the differences in training behavior between clean samples and noise samples. At the same time, in order to amplify the difference in loss between samples, a relative noise weight estimation method based on KLIEPRPD was proposed to estimate the relative probability density of each sample and used it as the weight of the sample loss behavior. This method effectively recovered the performance of the malicious DoH traffic detection model after cleaning the CICDoHBrw2020 dataset. Experiments verified that this method had good stability and outperformed other noise detection methods.

Key words: noisy label detection, noise weight, encrypt malicious traffic, DoH traffic, differential training