Abstract: solemnly declare: See the original title, if infringement, please contact the author, will be revoked release! The following is a key part of this excerpt translation, please refer to the original.
Abstract
Federal learning machine learning method is a promising, it is personalized using a distributed set of data from a plurality of nodes (e.g., mobile device) to improve performance while providing privacy for mobile users. In the federal study, the training data is widely distributed on mobile devices as users maintained. Central polymerization party mobile device to update the global model collected locally by the local training data updates from a mobile device to train global model in each iteration. However, no reliable data may be a mobile device (ie user) to upload, resulting in learning tasks in the federal fraud. The user may intentionally do not reliably updated, for example, data poisoning attacks, or unintentionally performed, for example, high-speed movement or low-quality restrictions caused by the energy data. Therefore, to find credible and reliable user becomes critical in the Federal learning tasks. This paper introduces the concept of reputation as a metric. On this basis, we propose a reliable user options for federal learning task. Union chain is used as a method to the center in order to achieve an effective reputation management for users, without the need to reject and tampering. By numerical analysis, it proved that this method can improve the reliability of mobile networks in the Federal learning task.