A Hybrid Approach for Threat Detection in Social Networks Using Graph Metrics and GANs

Abstract

Social networks have become a fertile ground for malicious activities such as fake profiles, coordinated misinformation, and phishing attacks. To address these threats, we propose a hybrid detection framework that combines graph-theoretical analysis with Generative Adversarial Networks (GANs). In our approach, user interactions are modeled as a graph, and structural features such as degree, centrality, and clustering coefficient are extracted using the NetworkX library. Malicious behavior is simulated through a GAN trained on synthetic and real user profiles, allowing the discriminator to detect abnormal patterns. A subgraph from a Twitter dataset was used to validate the system, and the classifier was able to distinguish suspicious users based on a probability threshold. The experimental results demonstrate the effectiveness of the proposed method, achieving an accuracy of 91.2% and highlighting the potential of combining topological insights with generative models for adaptive and scalable threat detection in social networks.