Shangyu Xie

Title: Shangyu Xie and Label Inference in Split Learning Defenses.

Introduction

Shangyu Xie is an inventor based in Los Angeles, California. His patent work addresses privacy risks in split learning, a method that trains a neural network across separate submodels.

Latest Patents

Xie is associated with 1 patent: “Label Inference in Split Learning Defenses.” The invention describes training a composite neural network divided into first and second submodels. It uses a noise-laden backpropagation gradient to train a surrogate submodel by optimizing a gradient-distance loss, then combines the first submodel with the trained surrogate to compute an updated dummy label and infer label information associated with the second submodel. The approach also describes adding noise to a label or a shared backpropagation gradient to help protect label information.

Career Highlights

The patent highlights Xie’s work on machine-learning privacy and defenses against label inference in split learning. Its techniques focus on reducing the risk that information shared during model training could reveal labels.

Collaborations

Xie’s listed collaborators include Jiankai Sun and Xin Yang.

Conclusion

Shangyu Xie’s patent presents methods for addressing label privacy risks in split learning through surrogate-model training and noise-based protections.