AgentShield-Real: A Unified Multi-Source Neural Framework for Real-Time Financial Fraud Detection Across Heterogeneous Transaction Domains.

Financial fraud is no longer a standalone institution problem— it evolves across borders, platforms, and currencies faster than any single fraud detection system can follow, damaging the global economy by $534 billion in 2025 (Nilson Report, Issue 1232, 2025). Existing fraud detection frameworks optimise for a single transaction domain and fail to capture cross-domain laundering patterns — creating a detection blind spot that sophisticated criminal networks deliberately exploit across multiple transaction domains. This paper presents AgentShield-Real, the first unified multi-source neural framework that simultaneously learns from three complementary public benchmark datasets — PaySim (2,770,409 mobile money transactions), IEEE-CIS (590,540 real-world e-commerce transactions with 432 features), and Elliptic++ (46,564 labelled Bitcoin transaction nodes) — within a single cross-domain detection pipeline. AgentShield-Real fuses all three datasets into a single heterogeneous graph pipeline. Entity schemas across three datasets are reconciled through type-specific input projections that map heterogeneous node types to a shared 128-dimensional embedding space. The framework uses focal loss to handle extreme class imbalance, a SHAP-grounded explanation layer aligned to EU AI Act Article 13 and SR 11-7 requirements, and delivers the first simultaneous multi-domain evaluation in the fraud detection literature. Experiments demonstrate that AgentShield-Real achieves F1 = 0.8459 and MCC = 0.8465 on PaySim — surpassing the best baseline by 37.6 F1 points — AUROC = 0.8802 with Recall = 0.7601 on IEEE-CIS, detecting 76% of all fraud events versus Gradient Boosting's 34.8%, and AUROC = 0.7212 on Elliptic++ under rigorous temporal split evaluation. Each dataset evaluated independently within the unified pipeline- the contribution is consistent multi-domain coverage, not end-to-end cross-chain tracking, which remains open research problem. No single baseline achieves operationally viable performance across all three domains simultaneously, establishing AgentShield-Real as the first consistent multi-domain fraud detection framework. All code, trained models, and datasets are publicly available to ensure full reproducibility.