Abstract: In the intelligent medical system, data from different medical institutions need to be shared across institutions. To ensure the privacy and security of data, federated learning is applied. However, in practical applications, federated learning still faces challenges such as data security, dynamic user management, and defense against malicious nodes. Based on this, this paper proposes a federated learning scheme based on privacy protection and anonymous tracing. This scheme uses homomorphic encryption and noise masking technology to encrypt the local model parameter information, ensuring the privacy of model parameters during the model update process. Then, proxy re-encryption is used, and the noise server generates and securely distributes the encrypted noise ciphertext, preventing the aggregation server from accessing the plaintext of noise and defending against gradient inversion attacks. At the same time, it can tolerate the situation where medical institutions fail to upload local model parameters, achieving efficient dropout-safe aggregation. Next, a traceable pseudo-anonymous identity authentication mechanism is adopted to protect the identities of medical institutions, and the changes in model accuracy are used to identify malicious nodes. When a medical institution is repeatedly marked as malicious, secret value sharing is used to trace the real identities of medical data. Finally, experimental simulations show that compared with existing schemes, the proposed scheme can effectively resist gradient inversion attacks, maintain model accuracy, and achieve efficient dropout-safe aggregation.

Key words: privacy protection, federated learning, homomorphic encryption, anonymous tracing, smart healthcare

摘要： 在智慧医疗系统中，需要不同医疗机构的数据进行跨机构共享，为了实现数据的隐私性与安全性，联邦学习被应用其中。然而，联邦学习在实际应用中仍然面临数据安全、动态用户管理和恶意节点防御等挑战。基于此，提出了一种基于隐私保护与匿名追踪的联邦学习方案。该方案利用同态加密与噪声掩码技术对本地模型参数信息进行加密，确保模型更新过程中模型参数的隐私性。利用代理重加密，由噪声服务器生成并安全地分发噪声密文，使得聚合服务器无法访问噪声明文，防止梯度逆推攻击，同时可以容忍医疗机构上传本地模型参数失败的情形，实现高效的掉线安全聚合。采用可追踪的伪匿名身份认证机制，实现对医疗机构的身份保护，并通过模型准确率变化来识别恶意节点，当某医疗机构多次被标记为恶意时，使用秘密值共享可追溯医疗数据的真实身份。实验模拟表明，与现有方案相比，所提方案可以有效抵抗梯度的反演攻击并保持模型精度，并且能够实现高效的掉线安全聚合。

关键词: 隐私保护, 联邦学习, 同态加密, 匿名追踪, 智慧医疗