GAN-TM:一种基于生成对抗网络的流量矩阵推断机制

GAN-TM:A Traffic Matrix Inference Mechanism Based on Generative Adversarial Network

随着当前互联网技术的快速发展,网络规模和复杂度不断提高,由于流量矩阵对于网络管理、流量工程、异常检测等都具有重要意义,因此准确测量流量矩阵对于计算机网络而言极其重要。当前针对流量矩阵的测量机制主要可以分为直接测量法和估计推断法,其中估计方法又包括简单统计反演法、附加链路测量信息法以及测量反演结合法。现有测量机制在准确性和测量耗费方面存在较多问题,直接测量的方法虽然可以保证准确性,但网络规模的扩张及网络结构的日趋复杂化使其在实现上存在困难,而流量矩阵推断问题在线性求解上固有的高度病态特性又使得估计推断法时常难以发挥作用,因此需要一种新的方法以更通用的方式解决现有问题。该文借鉴生成对抗网络(GAN)在图像恢复方面的作用,提出了一种基于生成对抗网络的流量矩阵推断机制GAN-TM。GAN-TM能够基于部分测量信息,建立起基于掩码矩阵评估的卷积生成对抗网络模型,利用部分测量信息对缺失的流量矩阵进行推断。实验结果表明,在数据缺失率低于30%的情况下,GAN-TM的推断误差能够控制在0.10以内。

With the rapid development of current Internet technology, the scale and complexity of network are increasing. Because traffic matrix is of great significance for network management, traffic engineering and anomaly detection, it is extremely important to accurately measure traffic matrix for computer network. At present, the measurement mechanism of traffic matrix can be divided into direct measurement method and estimation and inference method, and the estimation method includes simple statistical inversion method, additional link measurement information method and measurement inversion combination method. The existing measurement mechanism has many problems in accuracy and measurement cost. Although direct measurement method can guarantee the accuracy, but the expansion of network scale and the network structure of the increasingly complicated in the implementation difficulties, inference and traffic matrix in the height of the inherent in solving linear pathological features and makes estimation method was often difficult to play a role, so need a new kind of method in a more general way to solve the existing problems. In this paper, GAN-TM,a traffic matrix inference mechanism based on generative adversarial networks(GANs),is proposed based on the function of GANs in image restoration. GAN-TM establishes a convolutional generation adversation network model based on mask matrix evaluation, and uses part of the measurement information to infer the missing traffic matrix. The experiment shows that when the data missing rate is 30%,the inference error of GAN-TM can be controlled within 10%.