FraudBench Benchmarks Adversarial Robustness in Financial Risk Assessment

This research highlights a critical gap in evaluating the adversarial robustness of machine learning models used in financial fraud and credit-risk detection. Traditional robustness assessments often fail to account for the unique characteristics of financial tabular data, such as domain-specific constraints, severe class imbalance, and asymmetric attacker capabilities. The paper argues that robustness is not solely a model attribute but also heavily influenced by the evaluation protocol itself. To address this, the authors present FraudBench, a novel benchmark designed for protocol-sensitive adversarial robustness evaluation. FraudBench assesses models under three distinct protocols: unconstrained attacks, post-hoc feasibility filtering, and deployment-aware constraint-integrated attacks. Experiments across four public financial datasets and various model types (neural, tree-based, ensemble) reveal that robustness conclusions vary significantly based on the protocol. For instance, integrating constraints directly into attack generation yields vastly more feasible adversarial examples than post-hoc filtering, underscoring the necessity of protocol-aware evaluation for accurate risk assessment.