DRnet: Dynamic Retraining for Malicious Traffic Small-Sample Incremental Learning

Deep learning has achieved good classification results in the field of traffic recent years due to its feature representation ability. However, existing technology cannot meet requirements for incremental tasks online scenarios. In addition, high concealment and fast update speed malicious traffic, number labeled samples that can be captured is scarce, small drive neural network training, resulting poor performance model. Therefore, this paper proposes an method small-sample classification. The uses pruning strategy find redundant structure dynamically allocates neurons training based on proposed measurement according difficulty new class. This enables perform without excessively consuming storage computing resources, reasonable allocation improves accuracy classes. At same time, through knowledge transfer method, model reduce catastrophic forgetting old class, relieve pressure large parameters with data, improve performance. Experiments involving multiple datasets settings show our superior established baseline terms accuracy, 50% less memory.