When Deep Learning Meets Inter-Datacenter Optical Network Management: Advantages and Vulnerabilities

To realize cost-effective and adaptive network control and management (NC&M) on inter-datacenter optical networks (IDCONs), people have considered network virtualization to let the operator of an IDCON work as an infrastructure provider (InP), which can create virtual optical networks (VONs) over the IDCON for tenants. In this paper, we use this network scenario as the background, and try to integrate deep learning (DL) based traffic prediction in the NC&M of the IDCON and the VONs created over it. We first design the service provisioning framework in which each tenant uses a DL module to predict the traffic in its VON and will submit a VON reconfiguration request to the InP, when it sees a significant mismatch between future traffic and the allocated resources in its VON. Then, the InP will invoke the VON reconfiguration to make the VON be better prepared for future traffic. An adaptive and scalable DL-based traffic predictor is proposed together with a cognitive service provisioning algorithm to exploit the temporal and spatial characteristics of interDC traffic and achieve effective service provisioning based on precise and timely traffic prediction. Next, we consider the situation where a tenant leverages “machine-learning-as-a-service” and outsources the training of its DL module to a third-party entity for overcoming its resource limitations, and analyze the induced vulnerabilities due to data poisoning. Our simulation results indicate that with our proposal, the InP can invoke VON reconfigurations timely and improve the service provisioning performance of each VON significantly. Meanwhile, the results also demonstrate that our data poisoning scheme can easily bypass the normal validation of the DL module and generate significant adversarial effects.