<p>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 <InlineEquation ID="IEq1"> <EquationSource Format="TEX">$\varepsilon _l = 4000$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <msub> <mi>ε</mi> <mi>l</mi> </msub> <mo>=</mo> <mn>4000</mn> </mrow> </math></EquationSource> </InlineEquation>, AdaCOS improves model accuracy by an average of <InlineEquation ID="IEq2"> <EquationSource Format="TEX">$3.6\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>3.6</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> (std <InlineEquation ID="IEq3"> <EquationSource Format="TEX">$\pm 0.42\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mo>±</mo> <mn>0.42</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation>, <InlineEquation ID="IEq4"> <EquationSource Format="TEX">$95\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>95</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> <InlineEquation ID="IEq5"> <EquationSource Format="TEX">$\textrm{CI}$</EquationSource> <EquationSource Format="MATHML"><math> <mtext>CI</mtext> </math></EquationSource> </InlineEquation>) compared to fixed Top-<i>K</i> methods. In non-IID data scenarios, it reduces accuracy degradation by approximately 18<InlineEquation ID="IEq6"> <EquationSource Format="TEX">$\%$</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation>. Furthermore, to achieve the same privacy protection level, AdaCOS requires approximately 22<InlineEquation ID="IEq7"> <EquationSource Format="TEX">$\%$</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation> 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.</p>

AdaCOS: adaptive differential privacy shuffle model based on cosine similarity

期刊论文
出版模式：

开放获取
发表日期：

2026年1月6日

查看全文

Zi Ye,
Guangyao Wang,
Chao Ma,
Ziming Guo,
Hai Huang

摘要

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 $\varepsilon _l = 4000$ $ε_{l} = 4000$ , AdaCOS improves model accuracy by an average of $3.6\%$ $3.6 %$ (std $\pm 0.42\%$ $\pm 0.42 %$ , $95\%$ $95 %$ $\textrm{CI}$ $CI$ ) 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.