Anomalous Link Detection in Dynamically Evolving Scale-Free-Like Networked Systems

Networked systems are becoming increasingly complex and interconnected, making them more vulnerable to attacks. Identifying anomalous communication links is of paramount importance to ensure the integrity and security of a system. Existing methods based on static graphs only capture the interactions in networked systems using a single snapshot. Dynamically evolving graph-structured representations are gaining traction in modeling networked systems due to their ability to capture evolving relationships and complex interactions in time. In this study, we propose probabilistic approaches to predict future anomalous links in dynamically evolving scale-free-like networks. Specifically, we introduce three probabilistic models exploiting the scale-free property of networked systems to predict future link probabilities based on historical data. We conduct a grid search on the Receiver Operating Characteristic (ROC) curves of the models to determine the optimal decision boundaries for identifying anomalous links. We evaluate the performance of our proposed models on a synthetically generated dataset simulating a dynamically evolving real-world scale-free-like communication network. Our empirical results show that the proposed models' accuracy rates change between 63.20% to 96.50%, while the F1 score is between 66.25% and 96.52%. We also show that our approach can discern the direction of future links and estimate their probabilities based on link orientations.