基于粒子群优化的差分隐私深度学习模型

doi:10.19678/j.issn.1000-3428.0065590

摘要/Abstract

摘要：

在深度模型差分隐私保护中，过大的梯度扰动噪声会造成数据可用性下降。提出一种基于粒子群优化算法的深度学习模型，根据粒子群优化策略，将粒子的位置映射到网络参数，寻找个体历史最优位置和全局历史最优位置。对全局最优粒子位置所得梯度进行扰动重新参与模型训练，在不改变参数和梯度结构的前提下利用网络传播属性对噪声参数进行优化，运用带扰动的位置参数获取目标函数，最小化经验风险函数，降低噪声对模型效用的损害。在联邦学习模式下，本地客户端利用粒子位置参数更新出本地环境中个体和全局历史最优位置，找到当前最优参数加以扰动，再上传至中心服务器。当每个客户端再次训练时，从服务器请求获取优化、扰动和聚合后的广播参数，参与合作学习。实验结果表明，该算法模型在保护隐私的同时能更好地保留模型可用性，相比DP-SGD算法，在Mnist、Fashion-Mnist和Cifar-10数据集上的准确率分别提高了5.03%、9.23%和22.77%，能够加快联邦学习中的模型收敛，在独立同分布和非独立同分布数据上都具有较好的表现。

关键词: 差分隐私, 噪声参数, 粒子群优化, 网络训练, 联邦学习

Abstract:

Excessive noise from gradient perturbations can lead to usability degradation in deep models with differential privacy protection. A deep learning model with Differential Privacy(DP) based on the Particle Swarm Optimization(PSO) algorithm is proposed to address this issue. The algorithm maps particle positions to network parameters according to the PSO strategy, identifying individual and global historical optimal positions. These global optimal particle positions are then used as gradients, which are perturbed and incorporated into model training. By leveraging network propagation properties without altering the parameters and gradient structure, noise parameters are optimized to realize the objective function. The model is then updated using optimized parameters with perturbation, minimizing the empirical risk function and reducing the impact of noise on model utility. In federated learning, local clients update individual and global historical optimal positions using particle position parameters within their local environment. They identify current optimal parameters to be perturbed before uploading them to the central server. Clients then request optimized, perturbed, and aggregated broadcast parameters from the server for further training in this cooperative learning process. Experimental results demonstrate that the proposed algorithm model effectively preserves model utility while protecting privacy. Compared to the DP-SGD algorithm, the accuracy of Mnist, FashionMnist, and Cifar-10 datasets increases by 5.03%, 9.23%, and 22.77%, respectively.The proposed algorithm accelerates model convergence in federated learning and exhibits strong performance with IID and non-IID data.

Key words: Differential Privacy(DP), noisy parameters, Particle Swarm Optimization(PSO), network training, federated learning

张攀峰, 吴丹华, 董明刚. 基于粒子群优化的差分隐私深度学习模型[J]. 计算机工程, 2023, 49(9): 144-157.

Panfeng ZHANG, Danhua WU, Minggang DONG. Differential Privacy Deep Learning Model Based on Particle Swarm Optimization[J]. Computer Engineering, 2023, 49(9): 144-157.

https://www.ecice06.com/CN/Y2023/V49/I9/144

图/表 15

图1 本文方案的参数优化策略

Fig.1 Parameter optimization strategy of the proposed scheme

图2 本文模型在不同惯性权重下的准确率

Fig.2 Accuracy of the proposed model under different inertia weights

图3 本文模型在不同迭代数下的准确率

Fig.3 Accuracy of the proposed model under different iterations

图4 Mnist数据集在MLP网络中不同隐私水平下训练的测试准确率变化曲线

Fig.4 Test accuracy variation curve of Mnist dataset trained under different privacy levels on the MLP network

图5 Mnist数据集在CNN网络中不同隐私水平下训练的测试准确率变化曲线

Fig.5 Test accuracy variation curve of Mnist dataset trained under different privacy levels on the CNN network

图6 Fashion-Mnist数据集在MLP网络中不同隐私水平下训练的测试准确率变化曲线

Fig.6 Test accuracy variation curve of Fashion-Mnist dataset trained under different privacy levels on the MLP network

图7 Fashion-Mnist数据集在CNN网络中不同隐私水平下训练的测试准确率变化曲线

Fig.7 Test accuracy variation of Fashion-Mnist dataset trained under different privacy levels on the CNN network

图8 不同算法在Mnist数据集上的IID测试准确率

Fig.8 IID test accuracy of different algorithms on Mnist dataset

图9 不同算法在Mnist数据集上的non-IID测试准确率

Fig.9 non-IID test accuracy of different algorithms on Mnist dataset

图10 不同算法在Fashion-Mnist数据集上的IID测试准确率

Fig.10 IID test accuracy of different algorithms on Fashion-Mnist dataset

图11 不同算法在Fashion-Mnist数据集上的non-IID测试准确率

Fig.11 non-IID test accuracy of different algorithms on Fashion-Mnist dataset

参考文献 29

1	ZHANG Y J , BAI G D , ZHONG M Y , et al. Differentially private collaborative coupling learning for recommender systems. IEEE Intelligent Systems, 2021, 36 (1): 16- 24. doi: 10.1109/MIS.2020.3005930
2	WANG S P , LI J , WU G J , et al. Joint optimization of task offloading and resource allocation based on differential privacy in vehicular edge computing. IEEE Transactions on Computational Social Systems, 2022, 9 (1): 109- 119. doi: 10.1109/TCSS.2021.3074949
3	HAN L M, ZHAO Y, ZHAO J. POSTER: blockchain-based differential privacy cost management system[C]//Proceedings of the 15th ACM Conference on Computer and Communications Security. New York, USA: ACM Perss, 2020: 925-927.
4	GUPTA R , ALAM T . Survey on federated-learning approaches in distributed environment. Wireless Personal Communications, 2022, 125 (2): 1631- 1652. doi: 10.1007/s11277-022-09624-y
5	朱黎明, 丁晓波, 龚国强. 图数据连续发布中的隐私保护方法. 计算机工程, 2022, 48 (5): 154- 161. URL
	ZHU L M , DING X B , GONG G Q . Privacy protection method in continuous publishing of graph data. Computer Engineering, 2022, 48 (5): 154- 161. URL
6	刘宇涵, 陈红, 刘艺璇, 等. 图数据上的隐私攻击与防御技术. 计算机学报, 2022, 45 (4): 702- 734. URL
	LIU Y X , CHEN H , LIU Y X , et al. State-of-the-art privacy attacks and defenses on graphs. Chinese Journal of Computers, 2022, 45 (4): 702- 734. URL
7	ABADI M, CHU A, GOODFELLOW I, et al. Deep learning with differential privacy[C]//Proceedings of 2016 ACM SIGSAC Conference on Computer and Communications Security. New York, USA: ACM Press, 2016: 308-318.
8	YU D, ZHANG H S, CHEN W, et al. Gradient perturbation is underrated for differentially private convex optimization[C]//Proceedings of the 29th International Joint Conference on Artificial Intelligence. New York, USA: ACM Press, 2021: 3117-3123.
9	BASSILY R, SMITH A, THAKURTA A. Private empirical risk minimization: efficient algorithms and tight error bounds[C]//Proceedings of the 55th IEEE Annual Symposium on Foundations of Computer Science. Washington D.C., USA: IEEE Press, 2014: 464-473.
10	ASI H, DUCHI J, FALLAH A, et al. Private adaptive gradient methods for convex optimization[EB/OL]. [2022-07-20]. https://arxiv.org/abs/2106.13756.
11	ANDREW G, THAKKAR O, MCMAHAN H B, et al. Differentially private learning with adaptive clipping[EB/OL]. [2022-07-20]. https://arxiv.org/abs/1905.03871.
12	PICHAPATI V, SURESH A T, YU F X, et al. AdaCliP: adaptive clipping for private SGD[EB/OL]. [2022-07-20]. https://arxiv.org/abs/1908.07643.
13	MOHAPATRA S , SASY S , HE X , et al. The role of adaptive optimizers for honest private hyperparameter selection. Artificial Intelligence, 2022, 36 (7): 7806- 7813.
14	PAPERNOT N, STEINKE T. Hyperparameter tuning with renyi differential privacy[EB/OL]. [2022-07-20]. https://arxiv.org/abs/2110.03620.
15	LI X, TRAMÈR F, LIANG P, et al. Large language models can be strong differentially private learners[EB/OL]. [2022-07-20]. https://arxiv.org/abs/2110.05679.
16	YU D, ZHANG H S, CHEN W, et al. Do not let privacy overbill utility: gradient embedding perturbation for private learning[EB/OL]. [2022-07-20]. https://arxiv.org/abs/2102.12677.
17	YU D, ZHANG H S, CHEN W, et al. Large scale private learning via low-rank reparametrization[EB/OL]. [2022-07-20]. https://arxiv.org/abs/2106.09352.
18	NASR M, SHOKRI R, HOUMANSADR A. Improving deep learning with differential privacy using gradient encoding and denoising[EB/OL]. [2022-07-20]. https://arxiv.org/abs/2007.11524.
19	GATI N J , YANG L T , FENG J , et al. Differentially private tensor train deep computation for Internet of multimedia things. ACM Transactions on Multimedia Computing, Communications, and Applications, 2020, 16 (3): 1- 20.
20	YANG X, ZHANG H, CHEN W, et al. Normalized/clipped SGD with perturbation for differentially private non-convex optimization[EB/OL]. [2022-07-20]. https://arxiv.org/abs/2206.13033.
21	HAN A, MISHRA B, JAWANPURIA P, et al. Differentially private Riemannian optimization[EB/OL]. [2022-07-20]. https://arxiv.org/abs/2205.09494.
22	DWORK C. Differential privacy: a survey of results[C]//Proceedings of International Conference on Theory and Applications of Models of Computation. Berlin, Germany: Springer, 2008: 1-19.
23	DWORK C , ROTH A . The algorithmic foundations of differential privacy. Foundations and Trends in Theoretical Computer Science, 2014, 9 (3/4): 211- 407.
24	KENNEDY J, EBERHART R. Particle swarm optimization[C]//Proceedings of International Conference on Neural Networks. Washington D.C., USA: IEEE Press, 2002: 1942-1948.
25	HOUSSEIN E H , GAD A G , HUSSAIN K , et al. Major advances in particle swarm optimization: theory, analysis, and application. Swarm and Evolutionary Computation, 2021, 63, 100868.
26	KAIROUZ P, OH S, VISWANATH P. The composition theorem for differential privacy[C]//Proceedings of the 32nd International Conference on Machine Learning. New York, USA: ACM Press, 2015: 1376-1385.
27	MCMAHAN H B, MOORE E, RAMAGE D, et al. Communication-efficient learning of deep networks from decentralized data[EB/OL]. [2022-07-20]. https://arxiv.org/abs/1602.05629.
28	GONG M G , XIE Y , PAN K , et al. A survey on differentially private machine learning. IEEE Computational Intelligence Magazine, 2020, 15 (2): 49- 64.
29	MIRONOV I. Rényi differential privacy[C]//Proceedings of the 30th IEEE Computer Security Foundations Symposium. Washington D.C., USA: IEEE Press, 2017: 263-275.

[1]	王景熠, 朱予涵, 周茹平, 刘耿耿. 基于动态粒子群优化的X结构Steiner最小树算法[J]. 计算机工程, 2024, 50(9): 226-234.
[2]	刘鹏欢, 周强, 王莹, 朱建锋, 罗宏杰, 王露, 王甜. 基于凹凸性和转向角的古陶瓷碎片二次匹配算法[J]. 计算机工程, 2024, 50(9): 356-366.
[3]	潘恩元, 钟原, 李平. 联邦异质性数据下半监督颈椎MRI分割模型[J]. 计算机工程, 2024, 50(9): 367-376.
[4]	李红娇, 王宝金, 王朝晖, 胡仁豪. 基于模型相似度与本地损失的双重客户端选择算法[J]. 计算机工程, 2024, 50(8): 153-164.
[5]	火久元, 王虹阳, 巨涛, 胡军. 多场景下基于AHP-EWM的人体健康状态评估模型研究[J]. 计算机工程, 2024, 50(7): 372-380.
[6]	顾永跟, 高凌轩, 吴小红, 陶杰. 非独立同分布下联邦半监督学习的数据分享研究[J]. 计算机工程, 2024, 50(6): 188-196.
[7]	周春雷, 宋继勐, 沈子奇, 余晗, 雷杰, 林兵. 数联网标识解析系统中的标识数据布局策略[J]. 计算机工程, 2024, 50(6): 311-320.
[8]	高家豪, 胡创业, 丁男, 刘战东. 智能网联汽车中联合驾驶风格的交通流数据有效性分析[J]. 计算机工程, 2024, 50(6): 367-376.
[9]	熊世强, 何道敬, 王振东, 杜润萌. 联邦学习及其安全与隐私保护研究综述[J]. 计算机工程, 2024, 50(5): 1-15.
[10]	顾永跟, 李国笑, 吴小红, 陶杰, 张艳琼. 预算约束下多任务联邦学习激励机制[J]. 计算机工程, 2024, 50(5): 149-157.
[11]	卢晓天, 朴春慧, 杨兴雨, 白英杰. 基于贝叶斯网络的差分隐私高维数据发布技术研究[J]. 计算机工程, 2024, 50(5): 167-181.
[12]	黄君泽, 吴文渊, 李轶, 石明全, 王正江. 面向动态公交的离散分层记忆粒子群优化算法[J]. 计算机工程, 2024, 50(4): 20-30.
[13]	张晓均, 李兴鹏, 唐伟, 郝云溥, 薛婧婷. 云-边融合的可验证隐私保护跨域联邦学习方案[J]. 计算机工程, 2024, 50(3): 148-155.
[14]	宋华伟, 李升起, 万方杰, 卫玉萍. 非独立同分布场景下的联邦学习优化方法[J]. 计算机工程, 2024, 50(3): 166-172.
[15]	刘少杰, 文斌, 王泽旭. 基于联邦学习的多技术融合数据交易方法[J]. 计算机工程, 2024, 50(3): 182-190.

选择文件类型/文献管理软件名称

选择包含的内容