Robust Multimodal Financial Fraud Detection via Missing-Modality-Aware Adaptive Fusion

Financial statement fraud detectors increasingly combine several evidence modalities, such as current statements, accrual ratios, multi-year trajectories, and peer benchmarks, yet in operational screening any of these channels may be unavailable for a given firm-year. Detectors trained on largely complete inputs then emit scores whose scale depends on which modalities happen to be observed, and the pooled ranking used for audit targeting deteriorates sharply. This paper presents MAAF, a missing-modality-aware adaptive fusion framework that (i) condenses each modality into a standardized evidence score produced by a modality expert, (ii) learns cross-modal evidence surrogates that infer the evidence of an absent modality from whichever modalities are observed, using dense distillation targets instead of scarce fraud labels, and (iii) combines observed and surrogate evidence through a fusion layer whose weights, surrogate trust, and calibration are conditioned on the availability pattern and trained with masked-pattern augmentation and cross-pattern consistency. On 146,045 U.S. firm-years (1990–2014) with AAER-based fraud labels and a strict temporal protocol, MAAF attains a test AUC of 0.720 under natural missingness and retains 0.711 when each modality is removed with probability 0.7, whereas the strongest deep fusion baseline falls to 0.653 and the boosted-tree detectors fall to 0.605 or lower. Ablations show that evidence surrogates and masked-pattern training account for most of the robustness gain.