AdaCOS: adaptive differential privacy shuffle model based on cosine similarity

This paper proposes AdaCOS, a novel adaptive differentially private randomization model for privacy protection in federated learning. AdaCOS addresses the issues of accuracy loss and inadequate parameter importance assessment in traditional SDP-FL models, which stem from fixed Top-K parameter selection strategies. The proposed model introduces a dynamic Top-K adjustment mechanism based on cosine similarity, marking the first incorporation of modified cosine similarity as a smooth indicator of training stability for adaptive privacy perturbation control. Additionally, a novel network pruning technique integrating weight magnitude and Hessian values is developed to achieve more precise quantification of parameter importance. Experiments on three real-world datasets—MNIST, CIFAR-10, and CIFAR-100—demonstrate that under the same privacy budget of $$\varepsilon _l = 4000$$, AdaCOS improves model accuracy by an average of $$3.6\%$$ (std $$\pm 0.42\%$$, $$95\%$$ $$\textrm{CI}$$) compared to fixed Top-K methods. In non-IID data scenarios, it reduces accuracy degradation by approximately 18$$\%$$. Furthermore, to achieve the same privacy protection level, AdaCOS requires approximately 22$$\%$$ less Laplace noise standard deviation, significantly mitigating the interference of noise injection with model convergence. This research provides a reliable solution with high performance and strong privacy guarantees for balancing privacy protection and model efficiency in practical federated learning applications.