FC2L: Federated Continual Graph Contrastive Learning for Evolving Cross-Bank Money Laundering Detection

Anti-money laundering (AML) detection must jointly address privacy-preserving collaboration and evolving laundering behavior, yet existing graph learning methods largely treat these separately. Financial data exhibits a hybrid partitioning structure: accounts are horizontally partitioned across banks, each holding the full attribute set for its own accounts, while the transaction party (TP) holds the transaction graph and transaction-level features, vertically separated from bank-held account features. Transaction labels are held separately by an active party (AP), and regulations (e.g., GDPR, GLBA, DPDP) prevent any party from pooling raw data. Compounding this, evolving laundering typologies make sequential graph learning prone to catastrophic forgetting (CF). Existing federated AML methods assume simplified partitioning and static graphs, while continual graph learning methods assume centralized data and rely on replay buffers, both ill-suited to privacy-sensitive financial applications. We propose FC2L, a federated continual learning framework for edge-level AML detection that unifies this hybrid data collaboration with memory-free continual graph learning. Banks train local autoencoders to produce account embeddings, shared with the TP, which builds edge embeddings via a Graph Convolutional Network (GCN) over the transaction graph and classifies them with an MLP using labels provided by the AP. The TP computes loss gradients over these embeddings and sends the aggregated gradient back to each bank to update its local encoder. Raw account and transaction data stay with their respective parties; only embeddings and gradients are exchanged. To mitigate CF in the GCN encoder, we introduce Column-wise Adaptive Balanced Orthogonal Projection (ACOP), a memory-free continual learning mechanism that constrains encoder gradient updates using an accumulated subspace of past-task edge embeddings. Unlike prior orthogonal-projection methods (e.g., AdaBOP), which apply a single global stability-plasticity coefficient across all feature directions, ACOP assigns each retained direction its own coefficient based on its correlation with the current task gradient: strongly correlated directions receive a small penalty $\lambda_s$ to recover plasticity, while weakly correlated directions receive a large penalty $\lambda_\ell$ to preserve stability. We theoretically show that ACOP achieves a favorable stability-plasticity tradeoff compared to existing methods. To stabilize the classifier, we design a Dual task-specific Supervised Contrastive Learning (DualCL), aligning edge embeddings with class-conditioned classifier prototypes to counter representation drift under heterogeneous, non-IID bank data. Regularization terms further anchor these prototypes and their decision scores to prior-task values, mitigating forgetting as new laundering typologies emerge. We evaluate FC2L on the AMLSim and IBM-AML benchmarks under a class-incremental edge classification setting with five sequentially arriving laundering typologies (fan-in, fan-out, gather-scatter, scatter-gather, and cycle). Across local epochs (2-20), embedding dimensions (64/128), and GCN depths (2-5), FC2L consistently outperforms federated, continual, and contrastive learning baselines, including the strongest baseline (DualCL+ AdaBOP). With 2 local epochs and 64-dimensional embeddings, FC2L achieves 0.827 average F1 and 0.129 average forgetting on AMLSim, improving F1 by 1.8\% and reducing forgetting by 11.6\% relative to AdaBOP. On the sparser, more class-imbalanced IBM-AML dataset, FC2L achieves 0.227 average F1 and 0.164 average forgetting, an 11.3\% relative F1 gain and 11.8\% forgetting reduction over the same baseline; the low absolute F1 reflects the extreme class imbalance of IBM-AML rather than a weaker relative improvement. Ablations confirm both ACOP and DualCL contribute independently to performance, and experiments across 3-100 clients, 5-50 communication rounds, and Dirichlet heterogeneity $\alpha \in \{0.1, 1, 30, 100\}$ show robustness with minimal computational overhead over baselines. Overall, FC2L provides a unified framework for continual AML detection under hybrid federated partitioning, combining distributed graph representation learning with memory-free forgetting mitigation for evolving laundering behavior.