4.1. Trade-offs Across Millisecond SLA Constraints and Detection Sensitivity
The critical synthesis of analytical architectures highlights fundamental tensions between relational graph topology and feature-level tabular resampling within high-volume UPI ecosystems. While tabular ensemble frameworks effectively mitigate extreme class imbalances by synthesising minority class instances to isolate critical predictors such as geo-location anomaly flags and historical fraudulent behaviour indicators (Hybrid GAN-RF Architecture, 2026), they systematically treat transactions as isolated events. Conversely, graph attention architectures demonstrate that capturing topological transaction neighbourhoods and uncovering money mule networks provides vital relational context while maintaining inference speeds under 20 milliseconds (Sentinel-UPI, 2026). Despite these computational advancements, a prominent research gap persists in harmonising multi-layered behavioural features with dynamic, streaming graph embeddings in live interbank switches. Existing literature predominantly evaluates isolated classifiers or independent relational networks, leaving the unified determinants of detection latency, false positive suppression, and classification precision under high-frequency panel dynamics insufficiently articulated. Furthermore, several methodological limitations restrict current scholarly and practical findings. First, existing empirical evaluations rely heavily on synthetic datasets or simulated benchmarks rather than comprehensive, unmasked transaction telemetry from national payment switches. Second, the computational overhead associated with updating dynamic graph neighbourhoods during peak transaction surges poses operational risks to strict millisecond service-level agreements. Finally, evolving adversarial obfuscation techniques and regulatory compliance mandates necessitate adaptive online model retraining, an imperative rarely addressed in static evaluation regimes. Overcoming these limitations is essential for establishing robust financial surveillance across real-time payment networks.