四川轻化工大学学报（自然科学版）2023，Vol.36Issue(6)：49-56,8.DOI:10.11863/j.suse.2023.06.07
改进EfficientNet图像分类的恶意流量检测模型
Improved Malicious Traffic Detection Model for EfficientNet Image Classification
摘要
Abstract
The malicious traffic detection models based on deep learning traffic image classification mostly focus on improving the accuracy by improving the learner,while the problem of image sample imbalance and poor feature extraction caused by the complexity of the image in the image dataset is ignored.For this issue,a malicious traffic detection model based on improved EfficientNet image classification has been proposed.Firstly,the SE(Squeeze-Excitation)module is replaced by the ECA(Efficient Channel Attention)module for the attention mechanism,to enhance the Internet to extract the effective features of the image.Then,by replacing the loss function in the last Softmax layer of the EfficientNet model,the model's ability of extracting image features is improved.Finally,the tolerance mechanism is added to the model.In the case of the model training accuracy has not been improved,the callback of the model learning rate can be made according to the indexes,and then the detection accuracy of the model can be improved.Experimental validation is carried out by using a public dataset of 913 malicious network traffic PCAP visualization images.The average accuracy of the test is 97.51% and the loss rate is 0.02%.Compared with the original model,the accuracy is improved by 1.93%,the loss rate is reduced by 0.07%,which shows that the improved EfficientNet model in this thesis has certain applicability and effectiveness.关键词
恶意流量/注意力机制/EfficientNet/ECA/准确率Key words
malicious traffic/attention mechanism/EfficientNet/efficient channel attention/accuracy分类
信息技术与安全科学引用本文复制引用
周子云,黄洪..改进EfficientNet图像分类的恶意流量检测模型[J].四川轻化工大学学报（自然科学版）,2023,36(6):49-56,8.基金项目
四川省科技计划项目(2020YFG0151) （2020YFG0151）
桥梁无损检测与工程计算四川省高校重点实验室开放基金项目(2022QYJ06) （2022QYJ06）
四川轻化工大学研究生创新基金项目(y2021091) （y2021091）
