FraudGuardian: Self-Supervised and Adversarial Learning for Robust Financial Fraud Detection

Financial fraud detection is challenged by severe class imbalance, evolving adversarial tactics, and the demand for explainable decisions. To address these issues, we propose FraudGuardian, a novel deep learning framework for robust and interpretable fraud detection. FraudGuardian integrates two synergistic learning mechanisms: Self-supervised Consistency Learning (SCL) captures intrinsic normal patterns at the local event level to improve sensitivity to subtle anomalies, while Adversarial Feature Mining (AFM) actively synthesizes challenging samples to learn a more generalized decision boundary. These components are dynamically balanced through an adaptive multi-task optimization scheme, effectively mitigating data imbalance. Extensive experiments on real-world financial transaction datasets show that FraudGuardian significantly outperforms state-of-the-art methods, achieving 97.9% AUC, 90.6% PR-AUC, and 86.2% F1-Score on a challenging credit card fraud dataset, representing a 3.1% PR-AUC improvement over the best baseline. Ablation studies validate the contribution of each component. Moreover, FraudGuardian demonstrates strong generalization in cross-dataset and cross-attack-type evaluations, with a 9.1% F1Score improvement over baselines when detecting novel fraud strategies. The framework also provides interpretability by highlighting suspicious local patterns, offering a powerful and generalizable solution for enhancing secure transaction systems.