| In the era of big data,with the rapid development of computer technology and artificial intelligence technology,more and more data are recorded,stored and used around the world.With the widespread application of the Internet,a large amount of data has promoted the progress of science and technology and brought convenience to people’s lives.With the development of data analysis technology and the introduction of various data mining algorithms,data dissemination and sharing are becoming more and more frequent.The wider dissemination of diverse data such as text,voice,and images poses a serious threat to people’s privacy,and data privacy protection has become an increasingly serious issue.How to preserve data availability to a certain extent while protecting data privacy is a research hotspot in the field of data release privacy protection.Therefore,this paper uses feature importance ranking to design a privacy protection algorithm,and performs privacy protection preprocessing on pre-release data.The main contributions include the following two aspects:1)K-anonymity is a common model for privacy protection of text data publishing,and Kanonymity feature selection is an effective way to balance data privacy and usability.Most of the existing K-anonymous feature selection methods are filtering methods,which sort individual features according to feature importance,and select the most important feature from the feature ranking to start searching for K-anonymous feature subsets.Because each feature does not affect the classifier output independently,this search strategy cannot guarantee that the final selected feature subset achieves a good balance between classification performance and privacy protection.The filtering method is independent of the classification algorithm and has a limited impact on the classification performance of the classifier,while another form of feature selection is the encapsulated feature selection method,which evaluates the pros and cons of feature subsets according to the performance of the subsequent model,which costs a lot of time.Therefore,in view of the limitation that the filtering method is independent of the classifier and the cost of wrappered search is high,this paper combines the characteristics of filtering and encapsulation,and designs a hybrid K-anonymous feature selection algorithm.As an evaluation criterion for feature selection,K-anonymity feature selection is classified as a hidden means in K-anonymity privacy protection.Based on the forward search strategy and K-anonymity condition,more K-anonymous feature candidate subsets are generated,and then the classification performance of the candidate subsets is evaluated according to the encapsulation classifier,so as to select the efficient and privacy-protecting feature subsets.2)Adversarial attack technology is often used in the privacy protection of image data release.There are existing privacy protection methods based on adversarial perturbation,which can generate new images that are indistinguishable to the naked eye.The availability is guaranteed,but it is highly dependent on the target model and cannot be used on a variety of models.Portable.Aiming at these problems,this paper proposes a transferable image privacy protection algorithm based on feature importance adversarial attack.The loss function is designed by using the feature importance index with better generalization and the random discarding operation of the disturbance is introduced in the disturbance generation process to improve the transferability.This method of applying adversarial perturbation to the image does not significantly distort the original image,and can well preserve the appearance characteristics of the original image.Experiments on two public image datasets prove that the image privacy protection algorithm based on feature importance adversarial attack proposed in this paper effectively improves the transfer ability between different models. |
