Dependency & environment conflicts across nodes — different hospitals used different Python, CUDA, and hardware setups, which initially caused compatibility errors during model synchronization. I resolved this by containerizing each client using Docker and standardizing the runtime environment.

Slow training on low-resource systems — some nodes had limited compute and memory, leading to bottlenecks in federated rounds.I optimized the model pipeline, added smaller batch sizes, and enabled partial participation to keep training efficient.

Ensuring privacy while sharing model updates — raw gradients risked information leakage in early experiments.I implemented secure aggregation and parameter encryption to protect sensitive data.

Communication dropouts during federated rounds — unstable network connections interrupted training in remote sites.I added checkpointing and retry mechanisms so nodes could recover and rejoin training without restarting.

Coordinating multi-center testing & debugging — tracing issues across distributed clients was challenging.I built structured logging and monitoring dashboards to visualize client status and training metrics.