A deep learning-oriented data sensitive attribute desensitization system and method

Abstract

The invention discloses a deep learning-oriented data sensitive attribute desensitization system and method. The system includes a feature extractor and a privacy confrontation training module; an input end of the feature extractor is connected to a training data set, and an output end is connected to a privacy confrontation training module; The feature extractor is composed of a convolutional neural network, which is the core module of training. It is trained by the data center and distributed to individual users for subsequent local data preprocessing after training. The privacy adversarial training module includes a proxy attack classifier. This scheme proposes that privacy adversarial training puts privacy attributes on the decision hyperplane in the feature space, so that attackers cannot infer, and proposes a conditional reconstruction module to ensure that other information except privacy attributes is preserved, which can be effectively applied to downstream tasks. A joint optimization strategy is proposed to balance data privacy and data availability, so that the two can achieve optimal results at the same time.

Description

technical field [0001] The invention relates to the field of artificial intelligence (AI) data privacy security, in particular to a deep learning-oriented data sensitive attribute desensitization system and method. Background technique [0002] In recent years, deep learning has shown excellent performance in many fields, such as image classification, speech recognition, natural language processing, etc. The success of deep learning can be attributed in part to the large-scale training data that many service providers collect from their users to train more accurate models. However, these collected training data often contain a large amount of private information, such as race, gender, age, etc., which can be easily acquired by pre-trained models, and this privacy may be further used for targeted advertising or even other malicious purposes. Behavior. In order to solve the problem of privacy leakage in data collection, researchers have proposed many privacy protection techn...

AI Technical Summary

Problems solved by technology

However, this feature-based method will still cause privacy leakage under the inference attack of deep learning. The attacker can use the model trained on the public data set to infer the private information in the original data from the uploaded features. At the same time, it is necessary to pre-define the downstream The task is difficult to generalize to the actual application scenario, that is, the service provider expects the collected data to have similar usability to the original data and not only valid for a specific task, so it is necessary to propose an effective privacy protection method to resist inference attacks At the same time, ensure the availability of data on subsequent tasks

Images