Large Language Model-Assisted Multimodal Federated Sensing for Compliance Risk Screening in Cross-Border Trade

With the continuous expansion of cross-border agricultural supply chains, reliable food-safety risk screening under decentralized data ownership, heterogeneous sensing equipment, and jurisdiction-dependent, non-stationary regulatory limits pose a fundamental challenge to intelligent inspection systems. To address these limitations, this paper proposes an LLM-assisted regulation-aware heterogeneous multimodal federated learning framework that establishes a three-tier decoupling architecture. First, an available-modality and quality-adaptive fusion mechanism dynamically weights RGB images, near-infrared (NIR) spectra, and environmental time-series while enforcing strict zero weighting on unavailable modalities and continuous suppression of degraded observations. Second, a regulation-independent state-prototype collaboration and personalized federated learning strategy is developed on a unified continuous concentration scale, enabling cross-institutional transfer of physical contamination representations while avoiding the use of jurisdiction-dependent compliance labels to define shared prototype groups and preserving client-specific adapters against non-IID drift. Third, offline large language model (LLM) parsing transforms unstructured multi-jurisdictional trade regulations into machine-executable rule bases, which are coupled with an uncertainty-aware risk calibration module to dynamically determine exceedance probabilities and reinspection recommendations without retraining front-end sensing encoders when regulatory standards shift. Experimental evaluations on a multisource maize and aflatoxin B1 screening dataset in a workstation-based federated simulation demonstrate that the proposed framework achieves a mean absolute error (MAE) of 0.054±0.002 and a root mean square error (RMSE) of 0.078±0.003, while elevating the area under the curve (AUC) and F1-score to 0.862±0.008 and 0.833±0.009, respectively, reducing the false negative rate (FNR) to 0.121±0.007 (p<0.001). Under severe client heterogeneity and missing-modality scenarios, the proposed approach maintains an AUC of 0.823±0.010 and an F1-score of 0.796±0.011. Ablation and ownership-level analyses assess robustness, calibration (expected calibration error of 0.032), and regulatory adaptation with fixed sensing encoders, providing an end-to-end, traceable decision-support paradigm for international agricultural trade screening.