Training method for improving robustness of multi-branch prediction single model in Internet of Things scene

Abstract

The invention discloses a training method for improving robustness of a multi-branch prediction single model in an Internet of Things scene, and the method comprises: segmenting training data under the structure of a cloud terminal, distributing the training data to terminals, forming an environment with data diversity, setting different optimization targets for each terminal, and improving the structural diversity. And each terminal updates the weight issued by the cloud by depending on the training data and the optimization target of the terminal, then the cloud recovers the weight of each terminal for aggregation, finally training is completed through multiple cycles, and the optimal weight of the multi-branch prediction single model is stored. According to the method, the structural diversity of the multi-branch prediction single model is successfully combined with the data diversity under the scene of the Internet of Things, so that the robustness of the multi-branch prediction single model is further improved, the classification capability of adversarial samples is improved, and the multi-branch prediction single model can safely and reliably complete a classification task.

Description

technical field [0001] The invention belongs to the field of computers and relates to a training method for a multi-branch prediction single model. In particular, it involves improving the robustness of single network models with high precision but difficult to correctly classify adversarial samples. The idea of ​​federated learning is used in it, which solves the problem that today's Internet of Things requires safe and reliable artificial intelligence in scenarios suitable for the Internet of Things. . Background technique [0002] In recent years, the number of Internet of Things has been increasing, and at the same time, it has been more closely integrated with deep learning, especially in image processing tasks. However, these high-precision image classifiers are vulnerable to adversarial examples. Adversarial examples are intentionally added micro-perturbations on clean images. These perturbations will not interfere with the judgment of the human eye, but can fool th...

AI Technical Summary

Problems solved by technology

However, conventional training methods can only effectively utilize the diversity of model structures, and cannot reflect the diversity of data brought about by the continuous expansion of the Internet of Things into training.

Images