Assessing the Effects of Physical Layer Attacks on Content Accessibility and Latency in Optical CDNs

Content Delivery Networks (CDNs) are a major enabler of large-scale content distribution for Internetapplications. Many of these applications require high bandwidth and low latency for a satisfactory user experience,e.g, cloud gaming, augmented reality, tactile Internet and vehicular communications [1]. Replication is one of themost prominent solutions to meet the requirements of latency-sensitive applications [1, 2]. However, infrastructuredisruptions can greatly degrade the performance of such applications, or even cease their proper execution. Theextent of degradation can be exacerbated by malicious attackers that target the critical elements of the CDNphysical infrastructure to disconnect or severely degrade services.In this work, we assess the effects of physical-layer attacks performed by cutting optical fiber links on contentaccessibility and latency in CDNs. We perform preliminary experiments on the Germany50 network with 50 nodesand 88 links, considering the scenario where attackers cut the links with highest importance, i.e., betweennesscentrality (denoted by the number of shortest paths that traverse a link), in order to increase the effectiveness ofthe attack. By cutting links that are traversed by the majority of shortest paths, the attack forces a larger part ofservices to use longer paths, incurring higher latency. We consider the cases where each content has between 1and 4 replicas placed at the network nodes with highest closeness centrality, i.e., nodes closest to all other nodes.We evaluate the attack scenarios using the average content accessibility (ACA) [2] and the total replica latency.The ACA denotes the portion of nodes able to access any replica out of all nodes. To obtain the total replicalatency, we summarize the propagation latency in the fiber and the switching latency at the network nodes alongthe paths connecting each node to the closest replica. The propagation latency is based on the average path lengthand light propagation speed in fiber equal to 2x10 m/s. The switching latency is based on the number of hops, andfor simplicity we assume that each node adds 20 μs delay, which is supported by commercially available switches.Fig. 1a shows that the network maintains content accessibility for all nodes under attacks cutting up to 20% ofthe links. For more than 20% of the links cut, several nodes become unable to connect to a replica, even when 4replicas are placed in the network. The latency results presented in Fig. 1b-1d focus on the case when up to 20%of links are cut and all nodes still have accessibility to content. In the single replica case, cutting 20% of the linksdoubles the average distance to replica (Fig. 1b), the average number of hops (Fig. 1c), and the total latency toreplica (Fig. 1d). For 2 and 3 replicas, the latency degrades by 38% and 48%, respectively, while with 4 replicasthe latency degradation amounts to 34%. Given that latency-sensitive applications can require latency in the orderof 20 ms [1], propagation and switching latency account for 10% of the maximum latency under normal networkworking conditions. Under attacks, such latency can account for up to 20% of the maximum allowed value.Considering cloud processing and other factors that also contribute to the overall latency, current networks areoperating at the limit of the latency requirements, and the increase caused by attacks can be the tipping factorcausing the network to fail to meet the latency requirements. (a) (b) (c) (d)Figure 1. Results for Germany50 network topology for (1) to (4) replicas: (a) the average content accessibility (ACA); (b)the shortest distance to replica and the corresponding propagation latency; (c) the number of hops to replica and thecorresponding node latency; and (d) the total latency to replica.