DKFraudNet: a knowledge-guided adversarial learning framework for fraud user detection

Fraud user identification in telecommunications is hindered by scarce, noisy, and imbalanced labels, while expert rules may provide ambiguous or contradictory evidence.

We propose DKFraudNet, a knowledge-guided framework that integrates domain knowledge regularization, an attention-adaptive conditional generative adversarial network, and virtual category learning. Expert rules are organized into a Deterministic-Ambiguous-Contradictory evidence taxonomy for controlled pseudo-labeling. Class-conditioned augmentation alleviates data imbalance, and kernel-based similarity refines ambiguous samples. The framework was evaluated using two real-world city-level telecommunications datasets containing 59,015 and 52,183 users, respectively.

DKFraudNet consistently improved downstream classification performance under weak supervision. On the independent City 2 validation set, the CatBoost instantiation achieved an accuracy of 0.911 and an F1-score of 0.909. For operational review prioritization, DKFraudNet-XGBoost required reviewing 40.2% of unverified users to capture 80% of fraud users, corresponding to a 49.71% workload reduction relative to random review, with an AUPRC of 0.954. In operator-side blind verification, genuine-member identification and false-member screening achieved accuracies of 98.9% and 93.3%, respectively.

The results show that combining structured domain evidence, adaptive generative augmentation, and uncertainty-aware refinement improves robustness, data efficiency, and operational usefulness for fraud detection under weak supervision.