Work

Multi-Strategy Adversarial Continual Learning for Dynamic Financial Graphs

Year: 2026
Type: article
Abstract: Financial interaction data arrive as a continuous stream in which both the population and the behaviour of malicious actors change over time. A detector deployed on such a stream is pressed from two sides: retraining on recent activity erases knowledge of earlier fraud topologies, and the accounts being scored rewrite their own neighbourhoods in order to cross the decision boundary. Current solutions treat these pressures separately, and graph adversarial training is almost always carried out against a single family of perturbations. This paper models the interaction as a repeated allocation game played over strategy portfolios on both sides and introduces Multi-Strategy Adversarial Continual Learning (MSACL). On the attack side, five evasion operators, namely reputation camouflage, gradient-guided injection, attribute manipulation, temporal restructuring and degree dilution, compete for a bounded budget that is reallocated online by a multiplicative-weights rule driven by the damage each operator actually inflicts. On the defence side, the replay buffer distributes its slots over four complementary exemplar criteria whose shares follow the same rule, and replayed exemplars carry logit targets that anchor previously learned decision boundaries. On two real cryptocurrency trust networks and a synthetic anti-money-laundering stream, MSACL improves the average AUC under attack by 26.3 and 19.6 percentage points over the strongest baseline on the two real graphs, matches offline joint training on clean data, and keeps the forgetting rate between 1.2 and 3.6 points. A leave-one-strategy-out study shows that robustness acquired against one operator does not carry over to the others, which quantifies the cost of single-family adversarial training.
Language: English
Open Access status: diamond
OpenAlex ID: W7222912497
FWCI:
Cites: 40
Cited by:
Related to: 10