SUM Framework Boosts Federated Class Incremental Learning.

Real-world intelligent systems often require both distributed collaboration (Federated Learning, FL) and continuous adaptation to new tasks (Continual Learning, CL). Combining these, known as Federated Class Incremental Learning (FCIL), introduces complex "spatial-temporal catastrophic forgetting" due to interference from diverse clients and sequential tasks. Existing solutions typically tackle these interferences separately, often adding computational or communication burdens on client devices. This research proposes SUM (Unified Geometric Surgery on Spatio-Temporal Adaptation Vectors), a novel server-side framework. SUM reinterprets FCIL as a unified multi-task learning problem, representing both client and task updates as adaptation vectors in a shared parameter space. The core innovation lies in performing "geometric surgery" on these adaptation vectors during the aggregation phase on the server. Spatial SUM mitigates client-level interference within each training round, while causal online temporal SUM addresses cross-task interference over time. Crucially, SUM achieves this without requiring additional client-side computation, communication, or memory beyond standard federated training. Empirical results show SUM improving performance by up to 22% over prior FCIL methods across various benchmarks, demonstrating robustness and efficiency.