基于蜂群学习的医疗数据协作方法研究  

Medical data collaboration method based on swarm learning

作　　者：康海燕[1] 王嘉康 苏静茹 KANG Haiyan;WANG Jiakang;SU Jingru(School of Information Management,Beijing Information Science and Technology University,Beijing 100192,P.R.China;China Electronics Standardization Institute,Beijing 100007,P.R.China)

机构地区：[1]北京信息科技大学信息管理学院,北京100192 [2]中国电子技术标准化研究院,北京100007

出　　处：《重庆邮电大学学报(自然科学版)》2025年第2期282-294,共13页Journal of Chongqing University of Posts and Telecommunications(Natural Science Edition)

基　　金：国家社科基金项目(21BTQ079);教育部人文社科基金项目(20YJAZH046);未来区块链与隐私计算高精尖中心基金项目(GJJ-23)。

摘　　要：智慧医疗的发展推动了医疗数据的大规模应用,现有的医疗数据协作方法面临着数据隐私保护、模型性能优化和跨医疗机构间协作积极性差等问题,迫切需要革新与优化。提出基于蜂群学习的医疗数据协作方法(medical data collaboration based on swarm learning,MDC-SL),促进智慧医疗的发展;提出基于医疗数据定价的激励机制算法,根据参与方贡献度和本地模型质量进行定价,设置奖励币奖惩机制,以鼓励协作训练,可以使系统达到最优效用;提出蜂群学习动态模型聚合算法,通过控制聚合权重,使高质量模型在聚合中的贡献占比提高,可以提升训练模型的性能以及抵御投毒攻击的能力。此外,设计了基于本地化差分隐私的蜂群学习来防止模型参数传递过程中的数据泄露,增强医疗数据隐私保护。通过在医学数据集上进行实验,结果表明,该方法的模型性能比原蜂群学习更优,加入噪声导致的模型损失在预期范围之内,系统的平均奖励高于基线4.3百分点,从而验证了上述方法的有效性。The development of smart healthcare has promoted the extensive application of medical data.However,existing methods for medical data collaboration face challenges such as data privacy protection,model performance optimization,and poor motivation for collaboration across medical institutions,demanding innovation and optimization.To address these issues,we propose a medical data collaboration based on swarm learning(MDC-SL),thereby facilitating the advancement of smart healthcare.Firstly,an incentive mechanism algorithm based on medical data pricing is introduced,which prices contributions from participants and local model quality,utilizing a reward coin incentive mechanism to encourage collaborative training,thus optimizing system utility.Secondly,a dynamic model aggregation algorithm based on swarm learning is proposed to enhance training model performance and defend against poisoning attacks by controlling aggregation weights to increase the contribution of high-quality models in aggregation.Furthermore,a swarm learning approach based on localized differential privacy is designed to prevent data leakage during model parameter transmission,strengthening medical data privacy protection.Finally,experiments conducted on medical datasets demonstrate that the proposed method outperforms original swarm learning in model performance,with model losses due to added noise falling within expected ranges.Moreover,the average reward of the system exceeds the baseline by 4.3 percentage points,validating the effectiveness of the proposed methods.

关 键 词：蜂群学习 激励机制 差分隐私 医学数据协作 智慧医疗 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]