Privacy-Preserving Federated Land-Cover Segmentation Under Non-IID Data: Region-Aware Contrastive Regularization and DP-SGD

Federated learning supports collaborative satellite-image segmentation without centralizing source imagery, but non-IID data distributions and information leakage through model updates remain important challenges. We evaluate centralized U-Net, vanilla FedAvg, client-local region-aware contrastive FedAvg, client-side DP-SGD, and DP-SGD with fixed-prototype contrastive regularization on the DeepGlobe dataset using a fixed five-client Dirichlet partition (α = 0.5). Non-private contrastive FedAvg achieved the highest observed held-out test performance, with mIoU = 0.401 and pixel accuracy = 0.728, compared with mIoU = 0.352 for vanilla FedAvg. DP-SGD provided a formal patch-level privacy guarantee under the stated accounting assumptions but substantially reduced segmentation utility: standalone private training achieved mIoU = 0.138 at ε ≈ 4.82, while the matched private contrastive configuration achieved mIoU = 0.115 at the same privacy expenditure. Private training also incurred substantially greater computational costs. These single-run findings highlight a pronounced privacy–utility–computation trade-off and warrant validation across multiple seeds and client partitions.