基于内积加密的双向隐私保护医疗诊断云服务方案

Bilaterally Privacy-Preserving Medical Diagnosis Scheme with Functional Inner-Product Encryption

部署于云平台的医疗诊断服务不仅推进了医疗资源的整合,还提高了病情诊断的精准性和高效性,但是该场景用户失去了对个人信息的掌控,对高度敏感的病理数据来说安全与隐私保护是实现基于云平台决断的前提.云端医疗数据的隐私保护可以通过差分隐私、安全多方计算和同态加密等密码学技术实现,避免泄露用户医疗大数据中的隐私信息.差分隐私中引入随机噪声会降低计算精度,安全多方计算技术面临昂贵的通信成本,同态加密需要花费较大的时间加密深度学习模型.本文基于内积加密技术提出一种实现双向隐私保护的医疗诊断云服务方案,不仅保护用户医疗大数据中的个人隐私,而且帮助模型开发方避免云端部署造成的模型信息泄漏风险,甚至能降低隐私保护技术对医疗诊断效率和准确率造成的影响.为了保障用户个人隐私,方案中用户上传密文形态的个人医疗数据到云端服务器,云服务器通过密文数据预测疾病的结果.该方案的疾病诊断服务由云服务器提供并维护,而疾病诊断服务的实现依赖于模型开发方部署到云端的模型.模型开发方使用自行的深度学习算法训练明文形式的数据集,并获得预训练模型,使之可以处理密文形式的数据.实验分析表明,所述模型能完成CRC-VAL-HE-7K数据集上的结直肠癌诊断,与传统云端EfficientNet相比,本文方案仅损失少量的性能和响应效率.

Medical diagnostic services deployed on the cloud platform can promote the integration of medical resources and make the diagnosis of medical conditions more accurate and efficient.However,users may lose control of their personal information.Currently,cloud-based medical data privacy-preserving schemes can effectively protect users’privacy using cryptographic techniques such as differential privacy,secure multi-party computation,and homomorphic encryption.However,schemes using differential privacy or secure multi-party computation usually incur expensive communication costs,while homomorphic encryption-based solutions can hardly achieve efficient computation.More importantly,those schemes ignore the risk of model information disclosure.This paper proposes a privacy-preserving scheme for medical diagnosis based on inner-product encryption,which aims to protect the privacy of individuals in medical data,prevent the leakage of model information when model providers deploy models in the cloud,and improve the efficiency and accuracy of medical diagnosis.The proposed scheme allows users to upload their medical data to the cloud in ciphertext to ensure data confidentiality,and the server can determine the likelihood of diagnosing diseases with the data in ciphertext.The model for this prediction service is trained by the model provider using the unencrypted dataset and it is converted to a form that can handle ciphertext data after the training is completed.Moreover,the weights of the model are deployed to the cloud server in ciphertext form to provide online medical diagnosis service.The experimental analysis in this paper shows that the proposed model and framework exhibit practical effectiveness for colorectal cancer diagnosis tasks on the CRC-VAL-HE-7K dataset,with only a small performance degradation and latency increase compared to the traditional cloud-based EfficientNet.