FedCAD: Cross-modal semantic alignment and distillation for cross-domain heterogeneous federated learning

Recent advances in the IoT and edge intelligence have made the deployment of CLIP-style image-text models in edge-cloud architectures increasingly common for collaborative sensing. However, resource heterogeneity at the edge limits the feasibility of using a unified backbone across devices with different computational budgets. In addition, non-IID data and domain shifts can disrupt image-text alignment and cause semantic drift during federated aggregation. To address these challenges, we propose FedCAD, a federated learning framework for effective knowledge collaboration under cross-domain distribution shifts and heterogeneous edge environments through cross-modal semantic alignment and multi-level distillation. FedCAD consists of three main components: (i) a Latency-Distribution Co-aware Clustering (LDCC) strategy with heterogeneous model orchestration to alleviate resource disparities and straggler effects; (ii) a decouple-then-align dual-stage feature alignment mechanism that suppresses domain-specific noise and preserves the geometric consistency of the joint image-text embedding space, thereby enhancing cross-domain generalization; and (iii) a multi-stage collaborative distillation protocol spanning intra-cluster, inter-cluster, and cloud levels, which promotes cross-cluster semantic complementarity and cross-architecture knowledge fusion to mitigate knowledge fragmentation. Experimental results show that FedCAD consistently improves classification accuracy and target-domain generalization across multiple cross-domain classification benchmarks. In addition, results on Flickr30K and MSCOCO further demonstrate its effectiveness in image-text matching and cross-modal retrieval. Under the frozen-backbone and lightweight-adapter setting, only adapter parameters are transmitted during communication rounds, which to some extent supports its deployment applicability in bandwidth-constrained networks.