AdaCOS: adaptive differential privacy shuffle model based on cosine similarity

AdaCOS introduces a dynamic conductor — cosine similarity — to orchestrate privacy in federated learning. It rewards clients whose updates align with the global consensus by boosting their contribution allowance (raising K) and lowering noise, allowing cleaner signals to shine through. Meanwhile, it restricts clients with divergent updates by tightening their quota (lowering K) and applying stronger safeguards. This intelligent resource allocation, all within a fixed privacy budget, delivers stronger model accuracy under the same rigorous privacy guarantee. Springer