Adversarially Regularized Dynamic Graph Embeddings for Robust Community Evolution and Threat Detection in OSN

Abstract

Online social networks generate large-scale dynamic graph data, where existing systems often struggle with three issues: limited robustness under adversarial perturbations, lack of temporal consistency in evolving communities, and high energy consumption during continuous training. To address these limitations, an adversarially regularized and carbon-aware dynamic graph framework is proposed for threat detection in evolving networks. The method combines adversarially regularized dynamic embeddings (AR-DyGS), temporal community tracking, and a sequential threat profiling module. A carbon-aware training controller (CATC) is used to adjust embedding updates and snapshot processing to reduce computational load. Graph snapshots are normalized to [0, 1], filtered for low-activity edges (< 1%), and processed using embedding dimension d = 128, learning rate 0.001, batch size 512, dropout 0.25, and AdamW optimization. Adversarial testing uses bounded perturbations (≤ 3% edges, ≤ 1.5% nodes). Experiments are conducted on Twitter Higgs, Reddit Hyperlink Network, and DARPA SocialSim standard datasets. Results show consistent performance with around 95%–96% accuracy and 92%–93% F1-score across datasets. Energy consumption is reduced by about 40%–45% using CATC, while maintaining stable detection performance. Ablation studies show performance drops when individual modules are removed, confirming their contribution. The framework scales up to 450,000-node graphs with stable accuracy above 92%. The system is suitable for social media threat monitoring, fraud detection in interaction networks, and cybersecurity surveillance in dynamic environments. The approach maintains detection reliability while reducing computational cost and supporting scalable deployment in real-world online network systems.