Fraud learns too: continual graph learning under strategic adversarial drift in dynamic networks

Abstract

Financial transaction networks face a persistent threat from strategic adversarial drift, in which sophisticated actors manipulate graph structure to bypass detection. Conventional temporal graph neural networks tend to fail in this setting because they forget historical patterns when retrained and generalise poorly to novel structural perturbations. We address this gap with Game Theoretic Anticipatory Continual Graph Learning (GT-ACGL), a framework that casts fraud detection as a continuous Stackelberg game between a defender and an adaptive adversary. The framework combines three components: a bilevel anticipatory optimisation step that trains the defender against simulated future attacks, an Adversarial Motif Memory that retains topologically significant historical patterns without redundancy, and a predictive smoothing module that preserves temporal fidelity during high throughput batched training. We evaluate the approach on three large dynamic graph datasets. On the financial benchmark, Elliptic Temporal, GT-ACGL improves F1 by 11.0 percentage points over the strongest baseline under adaptive attack, with smaller but consistent gains on two behavioural interaction benchmarks. The framework also reduces the observed forgetting rate to below 6 percentage points and incurs only a $$1.45\times$$ training overhead relative to a standard temporal graph network. By modelling the cost of evasion inside a Stackelberg training objective, GT-ACGL encourages decision boundaries that remain comparatively stable under strategic structural perturbation. These results are empirical observations on the studied benchmarks, obtained against the specified edge addition threat model realised by our own attack generator. They are not guarantees of equilibrium behaviour, of the economic infeasibility of attack, or of robustness to the full range of real world fraud adaptations.