Federated Analytics Assisted Semantic Alignment for Secure and Privacy-preserving Image Classification

Federated learning (FL) offers a privacy-preserving paradigm for distributed target recognition, yet its efficacy is fundamentally challenged by hierarchical heterogeneity arising from multi-source imaging geometries, scattering mechanisms, and target categories. These physics-driven discrepancies not only degrade accuracy, but also amplify collaboration instability, induce cross-client parameter conflicts, and increase privacy risks of statistical exposure issues that generic FL methods cannot handle. To address these challenges, this paper presents FedCSA, a federated clustering framework with cross-client semantic alignment for security and privacy-aware federated recognition. Our approach integrates semantic-aware client clustering that groups clients with compatible scattering characteristics via probabilistic deep embedding and contrastive learning, mitigating negative transfer and improving robustness. We further introduce a federated-analytics-assisted class-conditional anchor alignment mechanism with secure aggregation, which constructs cluster level semantic representations while regularizing local feature distributions without exposing sensitive client statistics. Addition ally, a hierarchical parameter sharing strategy aggregates only shallow transferable layers while retaining client-specific deep representations, balancing collaboration efficiency, personalization, and privacy. Theoretical analysis establishes convergence guarantees under heterogeneity, while extensive experiments on widely-used public datasets including MSTAR and OpenSARShip demonstrate that FedCSA consistently outperforms state-of-the art federated baselines in accuracy, convergence stability, and communication efficiency, achieving a balanced trade-off among performance, privacy exposure control, and efficiency.