Differential privacy deep learning method based on feature region segmentation

Abstract

The invention relates to a differential privacy deep learning method based on feature region segmentation, and the method comprises the steps: obtaining the contribution degree of input features of different training data to model output through a layer-by-layer correlation propagation algorithm, and calculating the differential privacy contribution degree; the method comprises the following steps: adaptively segmenting an input feature into regions with different importance according to a difference privacy contribution degree by using a cohesive hierarchical clustering algorithm, adding Laplace noise to the input feature according to the importance degree of each region, and in order to prevent an attacker from obtaining sensitive data information through a data label, providing a new algorithm for the attacker to obtain sensitive data information. A loss function is converted into a polynomial form, Laplacian noise is added to coefficients of the polynomial, a deep learning model uses a differential privacy data training model, gradient is calculated through a noise loss function, and model parameters are updated for a subsequent training process; according to the method, good availability of the model can be maintained while the privacy of the deep learning model is ensured.

Description

technical field [0001] The invention relates to a differential privacy deep learning method based on feature area segmentation, belonging to the technical field of deep learning and network security. Background technique [0002] Deep learning combines low-level features to form more abstract high-level representation attribute categories or features by building multi-layer neural networks, so as to fully mine valuable information in data, and perform in various tasks such as image recognition, speech recognition, and natural language processing. Excellent. However, deep learning relies on massive data, which inevitably involves a large amount of private data; in addition, deep neural networks contain a large number of hidden layers, and the details of some individual data may be encoded into the model during the model training process. parameters, and even memorize the entire neural network, which brings unprecedented challenges to the data security and privacy of deep lea...

AI Technical Summary

Problems solved by technology

However, deep learning relies on massive amounts of data, which will inevitably involve a large amount of private data; in addition, deep neural networks contain a large number of hidden layers, and the details of some individual data may be encoded into the model during model training. parameters, and even memorize the entire neural network, all of which bring unprecedented challenges to the data security and privacy of deep learning

Images