Joint deep learning method capable of resisting generative adversarial network attacks

Abstract

The invention provides a joint deep learning method capable of resisting GAN (Generative Adversarial Network) attacks. The joint deep learning method comprises the following steps: initializing a deeplearning model; performing blind server initialization; and performing model protection joint learning and the like. By combining a matrix blinding technology and a random gradient descent method, blinding of input vectors and part of model parameters can be realized. According to the method, modeling and updating of an attacker local generative adversarial network are limited; meanwhile, the modes such as deep learning model right of use are limited, distributed trainees are allowed to utilize private data sets locally to train to obtain gradient update of model parameters, gradient update of each trainee is aggregated by a parameter server, and global update of a system model is achieved. According to the method, the GAN attack is resisted, the joint deep learning system model is protected, and the model accuracy and training data privacy protection requirements are greatly balanced.

Description

technical field [0001] The invention belongs to the fields of privacy data protection and deep learning, and specifically relates to a joint deep learning method capable of resisting generational confrontation network attacks. Background technique [0002] Joint deep learning refers to the use of cloud servers by multiple users to complete deep learning model training tasks under the premise of locally saving private training data. With the increasing demand for processing massive data, deep learning, as a machine learning method based on artificial neural network, has become more and more popular because of its powerful data feature learning ability, and has been widely used in computer vision, speech, etc. Among many practical problems such as recognition and natural language processing. Thanks to the high accuracy of classification and prediction results of various models, deep learning has now become the basis of Internet intelligent services. [0003] During deep lear...

AI Technical Summary

Problems solved by technology

However, the problem of model protection is also an important topic in the field of machine learning outsourcing computing. A complete and highly robust deep learning model often contains a large amount of training data information. If these training data contain sensitive information or private data, then for Misuse of deep learning models often leads to breach of privacy of training data

Images