How Malicious attack affects network’s functioning
We all know that networks surround our everyday life. People around us form networks; different cities connect with each other; web pages cover almost every topic link to one another. However, malicious attacks on large networks can sometimes yield unmeasured loss for everyone. For example, one terrorist attack on crucial transportation networks can potentially “paralyze airline traffic, electric power supply, ground transportation or Internet communication” (Schneider et al.). In order to figure out what kind of attacking strategy causes the most devastating result, researchers set up the experiment to test the robustness of the network under different attacks. The robustness of the network represents the resilience of a network under possible attacks. Node/link/community robustness measures the fraction of functional nodes/links/communities under attack.
The experiment is divided into two parts: testing the robustness under different attack strategies with or without protection. They also perform the attack on two different scales: small-scale node attacks and large-scale cluster attacks. There are six strategies in total:
- Random attack
- Attack nodes/clusters with the highest numbers of balanced edges in descending order
- Attack nodes/clusters with the highest influence (descending)
- Attack nodes/clusters with the highest degree/number of edges (descending)
- Attack nodes/clusters with highest influence (accounting the influence from weak ties) (descending)
- Attack nodes/clusters with the highest influence and recalculate the influence of each node every time (descending) (PageRank)
When the node or the cluster is under attack, their sign of edges reverses. For example, if A original has a positive relationship with B, it switches to a negative relationship after the attack.
When the system is under protection, 5% of the most influential nodes will not change their signs under any attack (except they choose random nodes for protection for the first strategy). Researchers calculate the system’s robustness when the k proportion of the nodes is under attack (k ranges from 0.1 to 1). The experiment uses two synthetic networks and four real-life networks to test the strategies.
There are several interesting results from the experiment:
- All networks are more vulnerable to cluster attacks than node attacks
- Strategies that attacks the highest influential nodes are the most successful, while the ones that attack nodes with the highest numbers of balanced edges do not work very well
- Real-life networks are more vulnerable than synthetic ones because they have more clusters
- Under small-scale node attacks, the robustness of the system reaches the lowest point when k=0.5(synthetic)/k=0.3(real) and slowly recovers when k keeps increasing
- The recovering robustness does not happen in cluster attacks
- The protected systems have much higher robustness than unprotected ones
I think the results make sense intuitively. For example, if individual nodes get attacked, they may still perform normally with a much higher distance. However, if an influential cluster is attacked, the whole system may be completely paralyzed. Also, we need to protect the highest influential nodes so that the effect of the attack will not spread out exponentially, similar to the small world phenomenon. I think the recovery of robustness with k>0.5 is the most interesting part. When more nodes get attacked, two nodes will become unbalanced first and then balanced again when both sides get attacked gradually. It’s a big problem when attackers only need to attack a small portion of the nodes and make about half of the system malfunction. I wonder if protect a cluster may do better than protect individual nodes. Develop better and cheaper protection mechanisms that help maintain the function of the system is important at this time.
Sources:
https://www.sciencedirect.com/science/article/pii/S0378437119321351
https://www.pnas.org/content/108/10/3838