基于拉格朗日对偶的小样本学习隐私保护和公平性约束方法  

Lagrangian Dual-based Privacy Protection and Fairness Constrained Method for Few-shotLearning

作　　者：王静红[1,2,3] 田长申 李昊康 王威[1,3] WANG Jinghong;TIAN Changshen;LI Haokang;WANG Wei(College of Computer and Cyber Security,Hebei Normal University,Shijiazhuang 050024,China;Hebei Key Laboratory of Network and Information Security,Hebei Normal University,Shijiazhuang 050024,China;Hebei Provincial Engineering Research Center for Supply Chain Big Data Analytics&Security,Hebei Normal University,Shijiazhuang 050024,China;Artificial Intelligence and Big Data College,Hebei University of Engineering Science,Shijiazhuang 050020,China)

机构地区：[1]河北师范大学计算机与网络空间安全学院,石家庄050024 [2]河北师范大学河北省网络与信息安全重点实验室,石家庄050024 [3]河北师范大学供应链大数据分析与数据安全河北省工程研究中心,石家庄050024 [4]河北工程技术学院人工智能与大数据学院,石家庄050020

出　　处：《计算机科学》2024年第7期405-412,共8页Computer Science

基　　金：河北省自然科学基金(F2021205014);河北省高等学校科学技术研究项目(ZD2022139);中央引导地方科技发展资金项目(226Z1808G);河北省归国人才资助项目(C20200340);河北师范大学博士基金项目(L2022B22)。

摘　　要：小样本学习旨在利用少量数据训练并大幅提升模型效用,为解决敏感数据在神经网络模型中的隐私与公平问题提供了重要方法。在小样本学习中,由于小样本数据集中往往包含某些敏感数据,并且这些敏感数据可能有歧视性,导致数据在神经网络模型的训练中存在隐私泄露的风险和公平性问题。此外,在许多领域中,由于隐私或安全等,数据很难或无法获取。同时在差分隐私模型中,噪声的引入不仅会导致模型效用的降低,也会引起模型公平性的失衡。针对这些挑战,提出了一种基于Rényi差分隐私过滤器的样本级自适应隐私过滤算法,利用Rényi差分隐私以实现对隐私损失的更精确计算。进一步,提出了一种基于拉格朗日对偶的隐私性和公平性约束算法,该算法通过引入拉格朗日方法,将差分隐私约束和公平性约束加到目标函数中,并引入拉格朗日乘子来平衡这些约束。利用拉格朗日乘子法将目标函数转化为对偶问题,从而实现同时优化隐私性和公平性,通过拉格朗日函数实现隐私性和公平性的平衡。实验结果证明,该方法既提升了模型性能,又保证了模型的隐私性和公平性。Few-shot learning aims to use a small amount of data for training and significantly improve model performance,and is an important approach to address privacy and fairness issues of sensitive data in neural network models.In few-shot learning,there is a risk of privacy and fairness issues in training neural network models due to the fact that small sample datasets often contain certain sensitive data,and that such sensitive data may be discriminatory.In addition,in many domains,data is difficult or impossible to access for reasons such as privacy or security.Also,in differential privacy models,the introduction of noise not only leads to a reduction in model utility,but also causes an imbalance in model fairness.To address these challenges,this paper proposes a sample-level adaptive privacy filtering algorithm based on the Rényi differential privacy filter to exploit Rényi differential privacy to achieve a more accurate calculation of privacy loss.Furthermore,it proposes a Lagrangian dual-based privacy and fairness constraint algorithm,which adds the differential privacy constraint and the fairness constraint to the objective function by introducing a Lagrangian method,and introduces a Lagrangian multiplier to balance these constraints.The Lagrangian multiplier method is used to transform the objective function into a pairwise problem,thus optimising both privacy and fairness,and achieving a balance between privacy and fairness through the Lagrangian function.It is shown that the proposed method improves the performance of the model while ensuring privacy and fairness of the model.

关 键 词：小样本学习 隐私与公平 Rényi差分隐私 公平性约束 拉格朗日对偶 

分 类 号：TP309.3[自动化与计算机技术—计算机系统结构]