基于模糊影响图的差分隐私算法保护效果评估方法
Assessment method on protection effectiveness of differential privacy algorithms based on fuzzy influence diagram
田月池 1李凤华 1周泽峻 1孙哲 2郭守坤 3牛犇3
作者信息
- 1. 中国科学院信息工程研究所,北京 100085;中国科学院大学网络空间安全学院,北京 100049;网络空间安全防御重点实验室,北京 100085
- 2. 广州大学网络空间安全学院,广东 广州 510006
- 3. 中国科学院信息工程研究所,北京 100085;网络空间安全防御重点实验室,北京 100085
- 折叠
摘要
针对隐私保护算法实际保护效果评估难的问题,提出了一种基于模糊影响图的差分隐私算法保护效果评估方法,实现对差分隐私算法的多维度评估,得出保护效果综合分数和等级.从算法安全性、算法可行性、隐私偏差性、数据可用性和用户体验5个方面出发,建立指标体系.使用模糊理论处理不确定性,通过影响图传递影响关系并计算该模糊影响图,得出保护效果分数和等级,据此反馈调整算法参数,实现迭代评估.提出正向化环节,解决截然相反的算法在某些情况下评估结果一样的问题.电-碳模型中的对比实验表明,所提方法能够对差分隐私算法的保护效果做出有效评价,消融实验进一步表明,正向化环节对算法的区分度起了关键作用.
Abstract
In response to the challenge of comprehensively assessing privacy-preserving algorithms,an assessment method on protection effectiveness of differential privacy algorithms based on fuzzy influence diagram was proposed,achieving a multi-perspective assessment of differential privacy algorithms with a comprehensive score and level as as-sessment results.Starting from five aspects—algorithm security,feasibility,privacy bias,data utility,and user experi-ence,an indicator system was established.Fuzzy theory was employed to handle uncertainties,while the diagram was used to propagate interactions between factors.The assessment score and level were obtained by calculating the fuzzy in-fluence diagram,and then used as feedback for parameter adjustment to achieve iterative assessment.Formalization link was proposed to solve the problem of completely opposite algorithms with idential evaluation results.Comparative ex-periments on electricity-carbon analysis model demonstrate the proposed method can assess the protection effectiveness of differential privacy algorithms effectively.Ablation experiments further show that the formalization link plays a key role in the discrimination of the algorithm.
关键词
隐私保护效果/综合评估/模糊影响图/差分隐私Key words
privacy protection effectiveness/comprehensive assessment/fuzzy influence diagram/differential privacy引用本文复制引用
基金项目
国家重点研发计划(2021YFB3100300)
国家自然科学基金(62332018)
国家社会科学基金重大项目(22&ZD147)
出版年
2024
国家科技期刊平台
NSTL
万方数据