Fed-DATrust: cross-stage risk coordination in task-subspace federated multimodal learning for reliable disaster assessment

Reliable disaster assessment is a complex distributed decision problem in which event heterogeneity, unstable modality quality, abnormal client behavior, and predictive miscalibration interact across the learning pipeline. This study proposes Fed-DATrust, a cross-stage risk-coordination framework for a three-class disaster damage assessment task derived from CrisisMMD. Shared image, text, and auxiliary-audio encoders remain fixed during federated adaptation, while task-specific projection, fusion, gating, reconstruction, branch, and classification modules are optimized and communicated. At each client, reconstruction error, gating uncertainty, predictive confidence, and classification margin are integrated into a sample-quality risk estimate that regulates dynamic adversarial training. At the server, client-clipped task-subspace updates are continuously reweighted according to graph isolation, local-loss deviation, and sample risk, then aggregated, server-clipped, and perturbed within the task subspace. During inference, temperature scaling calibrates class probabilities, while a decoupled composite anomaly score supports risk-based case prioritization without changing class predictions. Across three random seeds, Fed-DATrust achieved a Macro-F1 of 0.5829 and an expected calibration error (ECE) of 0.0601, improving over federated averaging by 0.0388 and 0.0335, respectively. It communicated 27.56% of the full-model parameter volume while slightly improving both measures relative to full-model aggregation. Stress tests further revealed improved robustness under missing vision and competitive aggregation under preset malicious-client conditions, while client-trust diagnostics showed that quality, loss, and update-geometry cues provide complementary evidence under event-level heterogeneity. These findings support task-subspace risk coordination as a practical design principle for reliable distributed multimodal intelligent systems.