With the emergence of AI technologies, intrinsic value of data is released and takes tremendous effects on numerous industries. In the context of regional medical care, data sharing and cooperating is in high demand, which can bring both financial and societal benefits. At present, however, medical data are locked inside medical facilities owing to legal risks and economic considerations. How to bring AI technologies into full play under this circumstance is a big challenge. In this paper, we propose Jupiter, an easy-to-use, secure, and high-performance platform for federated machine learning. Jupiter constructs a secure and highperformance aggregator cluster with SGX to efficiently aggregate the encrypted model parameters. Jupiter employs a stateful design to cooperate with medical facilities in regional medical systems with a fixed network connection. By providing an innovative programming abstraction, Jupiter makes model development more friendly to developers. The experiments show that with a low memory footprint, the throughput of a single node on an ordinary PC can reach 300 MB/s (with slice size fixed to 64 KB), and the aggregation primitive we built can process 11k aggregations per second.

随着人工智能技术的出现，数据的内在价值得到释放，对众多行业产生巨大影响。在区域医疗大背景下，数据共享与合作需求旺盛，既能带来经济效益，也能带来社会效益。然而，目前，出于法律风险和经济考虑，医疗数据被锁定在医疗设施内。如何在这种情况下充分发挥人工智能技术是一个很大的挑战。在本文中，我们提出了 Jupiter，一个易于使用、安全且高性能的联合机器学习平台。Jupiter 使用 SGX 构建了一个安全且高性能的聚合器集群，以高效聚合加密模型参数。Jupiter 采用有状态设计，通过固定网络连接与区域医疗系统中的医疗设施进行合作。通过提供创新的编程抽象，Jupiter 使模型开发对开发人员更加友好。实验表明，在内存占用低的情况下，普通PC上单个节点的吞吐量可以达到300 MB/s（切片大小固定为64 KB），我们构建的聚合原语每秒可以处理11k个聚合。