Graph Neural Network Based Anti‑Fraud Solution for Online Information Services

Recent rapid growth of China’s internet economy has led to massive black‑market fraud, evolving from simple system attacks to large‑scale exploitation of business risk‑control gaps, posing severe challenges to online service security.

58.com, as the country’s largest classified‑information platform, has built a security platform and anti‑fraud system; the authors first analyze black‑market fraud traits, summarize mainstream countermeasures, and identify three typical fraud patterns: aggregation, professionalism, and adversarial dynamics.

Common anti‑fraud techniques—credit‑list databases, expert rules, machine‑learning models, and relational networks—are described, highlighting their advantages and limitations, especially in the face of evolving fraud tactics.

The proposed solution applies graph neural networks (GNN) to capture deep, robust features of fraudulent users across three modalities: device aggregation, content similarity, and behavior co‑ordination. A user‑centric relational graph is constructed, weighting edges by device sharing, IP similarity (using a BM25‑inspired formula), content semantic similarity (via pretrained language model embeddings and angular cosine distance), and behavioral synchrony (hourly login/post vectors).

Feature extraction includes discretizing continuous attributes, selecting informative features via mutual information, and feeding the graph into the GraphSAINT sampling‑based GNN. GraphSAINT mitigates memory overflow by sub‑graph sampling, enables inductive learning, and reduces variance through influence‑aware node/edge sampling.

Experimental evaluation on a graph of 1.5 million nodes and 40 million edges (implemented with PyTorch Geometric) shows that the GraphSAINT‑based model improves precision and recall by 1–2 % over GraphSAGE, reduces training time by ~40 %, and demonstrates that content similarity contributes more discriminative power than behavior similarity in the information‑publishing scenario.

The authors conclude that multimodal graph construction and GNN‑based semi‑supervised learning significantly enhance fraud detection, and they outline future work: refining graph building, strengthening device‑fingerprint security, and continuously integrating emerging AI techniques into the anti‑fraud pipeline.