FedGAMMA Enables Federated Multimodal Graph Foundation Models

Multimodal-attributed graphs (MAGs), which integrate various data modalities like images and text with topological structures, are increasingly prevalent but often fragmented across privacy-restricted data silos. This fragmentation poses a significant challenge for learning broadly transferable models. To address this, FedGAMMA is proposed as a two-stage framework for federated multimodal graph foundation learning, focusing on semantic-structural alignment. The first stage, federated pre-training, involves a shared-private semantic enhancer that disentangles common cross-modal information from modality-specific details, aligning them via optimal transport. A topology-aware graph fusion module then decouples semantic and structural views using semantic residual graphs and dual positional encodings. Client similarity is estimated through a dual-channel affinity-aware aggregation mechanism without exposing raw data. The second stage, prompt-based fine-tuning, adapts the pre-trained encoder using lightweight graph-aware prompts, a shared prompt pool with controlled exploration, and channel-wise prompt synchronization. Experiments across twelve multimodal graph datasets show FedGAMMA consistently outperforming a wide range of baselines, with gains up to 12.96%, and also demonstrating superior performance in few-shot learning scenarios across multi-domain datasets.