Adversarial Training–Based Deep Imbalanced Learning
TL;DR
Extensive evaluations on seven real-world financial data sets demonstrate that ATDIL outperforms state-of-the-art imbalanced learning methods across multiple metrics while exhibiting superior resilience under adversarial conditions, offering a robust and practical framework for enhancing financial fraud detection systems.
Abstract
Financial fraud detection is crucial in the banking and financial sectors, but it faces considerable challenges because of class imbalance in transaction data and the growing threat of adversarial attacks. These issues frequently undermine the effectiveness of deep neural networks despite their demonstrated potential in this domain. To address these challenges, this paper proposes a novel approach, adversarial training–based deep imbalanced learning (ATDIL), which integrates imbalanced learning and adversarial defense into a unified approach. ATDIL leverages an adversarial autoencoder to efficiently synthesize high-quality minority-class samples that are informative and adversarial, while maintaining low computational complexity. The effectiveness of ATDIL is rigorously validated through both theoretical and experimental analyses. Theoretically, the optimal solution form for the inner optimization problem in ATDIL is derived, and its convergence under mild assumptions is established. Extensive evaluations on seven real-world financial data sets demonstrate that ATDIL outperforms state-of-the-art imbalanced learning methods across multiple metrics while exhibiting superior resilience under adversarial conditions. This combination of theoretical guarantees and empirical evidence highlights ATDIL’s ability to effectively address class imbalance and model security, offering a robust and practical framework for enhancing financial fraud detection systems. History: Accepted by Ram Ramesh, Area Editor for Data Science and Machine Learning. Funding: This work was supported by the National Natural Science Foundation of China [Grant 72401208], the Key Program of the National Natural Science Foundation of China [Grant 72331007], the Natural Science Foundation of Sichuan Province [Grant 2025NSFSC1981], the Postdoctoral Fellowship Program of the China Postdoctoral Science Foundation [Grant GZB20240504], the International Visiting Program for Excellent Young Scholars of Sichuan University (SCU), and the Humanities and Social Science Youth Foundation of the Ministry of Education of China [Grant 23YJCZH088]. Supplemental Material: The software that supports the findings of this study is available within the paper and its Supplemental Information ( https://pubsonline.informs.org/doi/suppl/10.1287/ijoc.2025.1251 ) as well as from the IJOC GitHub software repository ( https://github.com/INFORMSJoC/2025.1251 ). The complete IJOC Software and Data Repository is available at https://informsjoc.github.io/ .