The Impact of Source-Target Node Distance on Vicious Adversarial Attacks in Social Network Recommendation Systems

Social network recommendation systems play a crucial role in shaping user experiences by using graph neural networks and link prediction methods to tailor suggestions for new contacts. However, these systems are vulnerable to adversarial attacks orchestrated by malicious users using frameworks to manipulate recommendations artificially. In particular, frameworks such as SAVAGE have proven that conducting efficient and effective sparse vicious attacks on recommendation systems is possible. However, whether these adversarial systems remain effective and how much are impacted by the distance between the source and target nodes in the social network is unclear. This study investigates the impact of source-target path length on overall attack performance, revealing real-world applications such as social media friend/follower suggestions, where, for example, the source user may or may not share common connections with the target user and/or the friends of the target user. Analyzing a Twitter dataset, we found that SAVAGE outperforms other models, especially for longer source-target distances, with minimal impact on the network structure. These results make it highly relevant for real-world scenarios.