CICI: TCR: Transitioning Differentially Private Federated Learning to Enable Collaborative, Intelligent, Fair Skin Disease Diagnostics on Medical Imaging Cyberinfrastructure

新的隐私增强技术使人工智能（AI）和高级数据分析能够在敏感、分布式、孤立和异构数据上使用。这项工作旨在通过采用联邦学习（FL）和差分隐私（DP），将新的隐私和人工智能技术过渡到医学成像领域。在这种方式下，敏感的成像数据可以保存在本地，只有模型被共享和聚合，并有很强的隐私保证。实现的一个关键新组件将是更好地适应数据异质性，减少DP-FL中训练引起的偏差，并提高整体预测精度。这项工作将建立并促进皮肤病诊断的网络基础设施，而现实的部署和评估有望为其他也在敏感数据上运行的非医疗网络基础设施提供转化影响。本研究将不同的私人联邦学习（DP-FL）框架转变为医学成像网络基础设施（MICI），使莱姆病等皮肤病的协作、智能、公平诊断成为可能。从DP-FL的角度来看，关键的洞察力是一种精心设计的、与众不同的私有数据增强技术，它可以作为FL客户端的医学图像偏移。该偏移量与局部FL模型一起优化，降低了数据的异质性，包括DP引入的数据异质性，从而提高了诊断的准确性和公平性。从MICI的角度来看，该研究为不同的现实世界场景量身定制了DP-FL，从患者用例到医院用例。即研究将跨设备DP-FL转换为面向患者的智能皮肤病自我诊断移动应用，然后跨孤岛DP-FL转换为医院或医院部门参与的基于医院的诊断。同时，当同一患者在不同科室或医院存在不同特征时，垂直DP-FL也会发生过渡。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

New privacy enhancing technologies are enabling the use of Artificial Intelligence (AI) and advanced data analytics on sensitive, distributed, siloed, and heterogeneous data. This effort seeks to transition novel privacy and AI technologies to the domain of medical imaging by adopting Federated learning (FL) together with Differential Privacy (DP). In this fashion, sensitive imaging data can be kept local and only models are shared and aggregated with strong privacy guarantees. A key novel component of the implementation will be to better accommodate data heterogeneity, reduce training-induced bias in DP-FL, and improve overall prediction accuracy. The work will build and contribute to cyberinfrastructure for skin disease diagnosis, while the realistic deployment and evaluation promises to provide translational impact to other non-medical cyberinfrastructure also operating on sensitive data.This research transitions differentially private federated learning (DP-FL) framework to medical imaging cyberinfrastructure (MICI) with the enabling of collaborative, intelligent, fair diagnostics of skin diseases, such as Lyme diseases. From the DP-FL perspective, the key insight is a carefully crafted, differentially private data augmentation technique that serves as an offset to medical images at FL clients. The offset is optimized together with local FL models to mitigate data heterogeneity including those introduced by DP, thus improving diagnostics accuracy and fairness. From the MICI perspective, the research tailors DP-FL for different real-world scenarios ranging from use cases for patients to those for hospitals. That is, the research transitions cross-device DP-FL to patient-oriented mobile applications for intelligent self-diagnostics of skin diseases, and then cross-silo DP-FL to hospital-based diagnostics in which a hospital or hospital department participates. At the same time, vertical DP-FL is also transitioned when different features of the same patient exist across different departments or hospitals.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.