Graph Learning Against Adversaries: Bridging Structural Topology and Multimodal Semantics in Fraud Detection

Abstract

The digital economy's unprecedented expansion has been accompanied by increasingly sophisticated, networked fraudulent activities that exploit systemic vulnerabilities. Traditional tabular machine learning models evaluate entities in isolation, often failing to detect the collusive, multi-step strategies employed by modern adversaries. Consequently, Graph Neural Networks (GNNs) have emerged as the foundational architecture for advanced fraud detection, leveraging complex topological relationships to uncover illicit patterns. This paper systematically traces the evolutionary trajectory from general graph representation learning to Graph Anomaly Detection (GAD), and ultimately to the highly adversarial domain of fraud detection. We deconstruct the critical mathematical and structural challenges inherent in this domain, primarily extreme class imbalance and severe feature heterophily—a phenomenon where malicious entities actively deploy relational camouflage by linking to benign nodes, thereby neutralizing the smoothing mechanisms of standard homophily-assuming GNNs. We analyze the architectural adaptations developed to counteract these adversarial tactics. Furthermore, we critically re-evaluate recent advancements in the field, analyzing empirical evidence that questions the purported absolute superiority of Graph Transformers over meticulously optimized classic GNNs. Finally, we explore the emergent frontier of Large Language Model (LLM)-enhanced graph learning. We detail how sophisticated techniques, such as dual-granularity prompting and agentic semantic vectorization, are actively resolving the catastrophic information loss associated with early-stage multimodal feature encoding. By bridging discrete topological structures with continuous multimodal semantic reasoning, this paper charts the future trajectory of robust, scalable, and explainable adversarial graph learning.