Google recently announced a Federated computing program,A Chinese startups has made relevant results
Recently,Google launched a paper on Federated learning. Based on TensorFlow, the paper distributes the model training process distributedly on the mobile phone side, and aggregates the results into tens of millions of mobile models to aggregate and update to the cloud.The terminal computing power training model is implemented on the basis of guaranteeing user privacy and data ownership. At the same time,Google also announced the project — Federated computing,in the future will use a MapReduce-like computing framework to process log data,providing gengral-purpose computing service. Coincidentally,a Chinese startup has made a lot of relevant results.
Gravity.link is a Chinese startup that determined the strategic goals of shared computing in early 2018.Similar to Google’s Federated Computing (FL), it increases computational efficiency and guarantees data privacy and ownership by moving calculations to the edge. In May 2018, the team launched the first version of the MapReduce computing framework based on Android phones. At the end of 2018, in the test network of tens of thousands of nodes, the actual business processing operations were run.
Gravity-PMapReduce (PMR) and Google Co-training face similar issues such as limited terminal resources, stability, scale, security, and trustworthiness.
Google uses some free infrastructure, such as Safety Net, to ensure that the terminal performs legitimate programs in a jail-free environment (not expected to be available on domestic phones due to Google Play issues). The Secure Aggregation protocol is used to implement a model of an individual device that is only aggregated in the cloud and cannot be viewed. In the future, multi-tenant capabilities will be provided to enable individual devices to handle different tasks and optimize. In addition, by compressing data transmission, reducing network overhead and the like.
Google is currently only doing Federated training and can develop customized optimization solutions. Gravity’s PMR is a general-purpose data processing framework, so the solution will be different. PMR solves trusted and data security issues with TEE and uses TrustZone technology in ARM devices. The PMR is optimized based on the execution calculation of the calculation job, making the data transfer cost the lowest in the iterative calculation process. The underlying Gravity also provides a more flexible resource scheduling system, including resource management and networking. Includes a container-based VCU normalization scheme that provides multi-tenancy capabilities and a VPC networking solution that supports NAT penetration on public networks.
Regarding data privacy and ownership issues, only a combination of computing and storage can be completely resolved. Both FC and PMR ensure data compliance by moving data to the edge to avoid data collection to the cloud. However, there is a problem FC does not explain, the data is not transmitted to the cloud to represent the user’s terminal power and data? Are users willing to share these resources? Maybe Google will add terms in the user agreement in the future. Gravity’s model is different. Contributing computing resources and data is the user’s active behavior. Through a set of incentive modes, the user’s contribution will be recorded in the blockchain. These resources are provided to the entity as a cloud computing service (Gravity-GCloud), and users who provide computing power will periodically share revenue based on the deposit.
As early as April 18, Gravity was the first to launch the Android version of the phone, users can download and install, perform the corresponding tasks and gain revenue. Gravity currently supports not only Android phones but also other ARM devices and PC servers. It is working with multiple data centers and mines, and some early customers in the fields of Internet finance, digital marketing, e-commerce and genetic data are collaborating to prepare for commercialization.
Gravity provides a shared computing engine that combines idle mobile phones, terminals, PCs, and more into a huge computing cluster. Gravity provides a peer-to-peer MapReduce framework for processing big data, and P2P’s resource scheduling system makes efficient use of each node’s resources. The blockchain value network is used to enable resource sharers to obtain incentives in a network of mutual trust.