Scholarlink

<p>Deep learning on medical images classification intervention needs to use large data on multi-institutional datasets but privacy laws inhibit sharing of data (GDPR, HIPAA). Federated Learning (FL) facilitates collaborative training without data transfer; until now, the known methods can only address privacy, personalisation, and accuracy not at the same time in a multi-modal environment. We present MM-PFL-ADP, a framework that combines Vision Transformer (ViT) based multi-modal feature extraction in four new elements: (i) privacy budget allocation (independent of number of samples): Fisher information-based adaptive per-parameter privacy budget allocation (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\varepsilon _{\text {local}} / \varepsilon _{\text {shared}} = 1.5\)</EquationSource> </InlineEquation>); (ii) personalisation masks: dynamic KL divergence based personalisation masks; (iii) respect The framework gives formal client-level <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\((\varepsilon , \delta )\)</EquationSource> </InlineEquation>-DP guarantees on transmitted gradient updates, in <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(K = 10\)</EquationSource> </InlineEquation> simulated medical institutions. On the MRI-MS dataset, MM-PFL-ADP achieves <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(97.3\%\)</EquationSource> </InlineEquation> accuracy (95% CI: 96.9–<InlineEquation ID="IEq5"> <EquationSource Format="TEX">\(97.7\%\)</EquationSource> </InlineEquation>) at <InlineEquation ID="IEq6"> <EquationSource Format="TEX">\(\varepsilon = 1.5\)</EquationSource> </InlineEquation>, outperforming FedAvg (<InlineEquation ID="IEq7"> <EquationSource Format="TEX">\(92.1\%\)</EquationSource> </InlineEquation>) and DP-FedAvg (<InlineEquation ID="IEq8"> <EquationSource Format="TEX">\(87.3\%\)</EquationSource> </InlineEquation>) by large margins (<InlineEquation ID="IEq9"> <EquationSource Format="TEX">\(p &lt; 0.001\)</EquationSource> </InlineEquation>). The framework is <InlineEquation ID="IEq10"> <EquationSource Format="TEX">\(45\%\)</EquationSource> </InlineEquation> faster than FedAvg (47 vs. 85 rounds), has <InlineEquation ID="IEq11"> <EquationSource Format="TEX">\(47\%\)</EquationSource> </InlineEquation> less total communication and keeps <InlineEquation ID="IEq12"> <EquationSource Format="TEX">\(95.2\%\)</EquationSource> </InlineEquation> accuracy in case of extreme heterogeneity in data (<InlineEquation ID="IEq13"> <EquationSource Format="TEX">\(\alpha = 0.1\)</EquationSource> </InlineEquation>). The probability of membership inference attack has decreased to 52.1 which was close to the random baseline (<InlineEquation ID="IEq14"> <EquationSource Format="TEX">\(50\%\)</EquationSource> </InlineEquation>). MM-PFL-ADP shows that the concepts of privacy, personalisation, and accuracy are synergistic, but not oppositional to federated medical AI. The single-system Fisher information framework greatly simplifies the hyperparameter tuning problem and can meet formal privacy criteria. Before being deployed, prospective validation against the performance of expert radiologists is desired.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multi-modal personalized federated learning with adaptive differential privacy for medical image classification and a privacy-preserving approach

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月12日

查看全文
  • Adhi Siva M,
  • Chiranji Lal Chowdhary

摘要

Deep learning on medical images classification intervention needs to use large data on multi-institutional datasets but privacy laws inhibit sharing of data (GDPR, HIPAA). Federated Learning (FL) facilitates collaborative training without data transfer; until now, the known methods can only address privacy, personalisation, and accuracy not at the same time in a multi-modal environment. We present MM-PFL-ADP, a framework that combines Vision Transformer (ViT) based multi-modal feature extraction in four new elements: (i) privacy budget allocation (independent of number of samples): Fisher information-based adaptive per-parameter privacy budget allocation ( \(\varepsilon _{\text {local}} / \varepsilon _{\text {shared}} = 1.5\) ); (ii) personalisation masks: dynamic KL divergence based personalisation masks; (iii) respect The framework gives formal client-level \((\varepsilon , \delta )\) -DP guarantees on transmitted gradient updates, in \(K = 10\) simulated medical institutions. On the MRI-MS dataset, MM-PFL-ADP achieves \(97.3\%\) accuracy (95% CI: 96.9– \(97.7\%\) ) at \(\varepsilon = 1.5\) , outperforming FedAvg ( \(92.1\%\) ) and DP-FedAvg ( \(87.3\%\) ) by large margins ( \(p < 0.001\) ). The framework is \(45\%\) faster than FedAvg (47 vs. 85 rounds), has \(47\%\) less total communication and keeps \(95.2\%\) accuracy in case of extreme heterogeneity in data ( \(\alpha = 0.1\) ). The probability of membership inference attack has decreased to 52.1 which was close to the random baseline ( \(50\%\) ). MM-PFL-ADP shows that the concepts of privacy, personalisation, and accuracy are synergistic, but not oppositional to federated medical AI. The single-system Fisher information framework greatly simplifies the hyperparameter tuning problem and can meet formal privacy criteria. Before being deployed, prospective validation against the performance of expert radiologists is desired.