Detecting Anomalies in Communication Packet Streams Based on Generative Adversarial Networks

摘要

The fault diagnosis in a modern communication system is traditionally supposed to be difficult, or even impractical for a purely data-driven machine learning approach, for it is a humanmade system of intensive knowledge. A few labeled raw packet streams extracted from fault archive can hardly be sufficient to deduce the intricate logic of underlying protocols. In this paper, we supplement these limited samples with two inexhaustible data sources: the unlabeled records probed from a system in service, and the labeled data simulated in an emulation environment. To transfer their inherent knowledge to the target domain, we construct a directed information flow graph, whose nodes are neural network components consisting of two generators, three discriminators and one classifier, and whose every forward path represents a pair of adversarial optimization goals, in accord with the semi-supervised and transfer learning demands. The multi-headed network can be trained in an alternative approach, at each iteration of which we select one target to update the weights along the path upstream, and refresh the residual layer-wisely to all outputs downstream. The actual results show that it can achieve comparable accuracy on classifying Transmission Control Protocol (TCP) streams without deliberate expert features. The solution has relieved operation engineers from massive works of understanding and maintaining rules, and provided a quick solution independent of specific protocols.