Graph-Based Detection Research
RESEARCH · IN PROGRESS
Benchmarking graph-neural-network detection models against production-shaped abuse topologies to test whether structure-aware approaches outperform rule-based detection as adversaries adapt.
Question
Rule-based detection degrades as adversaries adapt; every rule teaches the next evasion. Can graph-neural-network approaches, which read structure instead of surface features, hold up against real-world coordinated-abuse topologies at platform scale?
Method
SliceNDice, BWGNN, and TGN are benchmarked against production-shaped synthetic topologies modeled on coordinated-abuse structures observed in the field. A research base of 200+ papers on graph-based fraud detection, indexed in a local knowledge-graph system, was narrowed to this candidate shortlist by match against real abuse graph shapes.
Status
Benchmarking in progress. Results land on this plate as they harden; the figure below is a placeholder for the comparison.