摘要
Abstract
In the task of detecting malicious encrypted traffic,the existence of noise tags seriously affects the generalization ability and detection accuracy of the model.To solve the above problems,a noise label learning method based on DC ASS(dual-confidence adaptive sample selection)is proposed to realize robust malicious encryption traffic detection.Firstly,the low dimensional features of samples are extracted by self encoder,and the feature confidence of samples is constructed.Then,the label confidence of samples is evaluated according to their performance in classification training.Finally,an adaptive selection threshold is proposed to select samples based on the dual confidence of feature space and label space,and filter noise samples dynamically to improve the robustness of the model.Experiments on CIRA-CIC-DoHBrw-2020 dataset show that the proposed method has good performance and stability in dealing with noise labels.The F1 scores of the method reach 86.686％,86.749％,83.199％respectively when the noise rate is 20％,30％,40％.Compared with the existing three methods,the method proposed in this paper shows the best performance under different noise rates,with the average performance improvement of 18.89％,37.34％,6.32％respectively.关键词
噪声标签学习/恶意加密流量检测/样本选择/深度学习/自编码器Key words
noise label learning/malicious encrypted traffic detection/sample selection/deep learning/autoencoder分类
信息技术与安全科学
