BLOSSOM: Block-wise Federated Learning Over Shared and Sparse Observed Modalities

BLOSSOM addresses a critical bottleneck in federated learning: real-world data heterogeneity where different edge devices (e.g., medical centers or autonomous vehicles) possess different subsets of sensors/modalities. While most FL research assumes uniform input shapes, BLOSSOM's block-wise approach allows for flexible training without discarding incomplete samples. The project currently has 0 stars but 4 forks within 3 days of release, which is a strong signal for a recently published research paper where peers are actively exploring the code. Defensibility is low (3) because it is a reference implementation of a paper; it lacks the ecosystem or software hardening of a tool like Flower or OpenFL. However, frontier risk is low because the major labs (OpenAI, Google) prioritize centralized training on massive scraped datasets, whereas this tool targets privacy-sensitive, siloed, and heterogeneous data environments. Its primary competitors are other research-grade FL frameworks like FedML or specific multimodal FL papers like FedMM. The displacement horizon is set to 1-2 years as the federated learning research field moves extremely fast, and new aggregation strategies for missing modalities are published frequently in venues like NeurIPS and ICML.