| With the rapid development of information technology and the popularization of the network,the data in the information system proliferate.In order to make effective use of data resources,big data technology emerges at the historic moment.By analyzing and using big data resources,it not only makes people’s lives more convenient,but also improves the efficiency of production and operation.To fully implement the big data strategy and release the technology and resource dividend has become a new driving force of economic growth in China.However,while enjoying convenience and efficiency,people also pay more and more attention to security threats such as information leakage and data abuse brought by big data.In traditional information security architecture,access control technology is the last barrier after virus protection,firewall and identity authentication.It ensures information security by restricting legitimate users to access system resources only within the authorized range.Under the condition of big data,because a large number of data resources or objects are unstructured data,its attribute extraction is difficult,and the security role of access control is limited.Image is a typical unstructured data,its information form is intuitive,the amount of data increases rapidly,semantic description has a high degree of abstraction and generality.For a long time,the security attributes of image data can only be marked manually,which is difficult to meet the demand of access control under the condition of big data.According to the characteristics of image data,this paper proposes a security attribute extraction method which is accurate,safe,scalable and has certain reasoning ability,and puts forward a series of technologies to provide attribute support for the access control of image data resources under the condition of big data.The following research results are obtained:1.Single image data security attribute generation is the foundation of image security attribute database.Aiming at the challenge of poor robustness of general-purpose image data security attribute generation due to open generation of image data and universal scene under the condition of big data,a single image security attribute generation technology based on edge detection and feature fusion is proposed.Taking advantage of the rich details of Sobel operator edge detection results and the intelligence of HED neural network detection,an intelligent edge detection algorithm combining Sobel operator and HED was proposed.By introducing image color reversal,edge nesting,broken line corrosion and other methods in the calculation process,dark line detection,edge position determination,scattered edge removal are realized and high precision edge image output is guaranteed.Based on the difference between general image features and edge image features,an image security attribute generation method based on feature fusion is proposed.A new branch is added into the method framework of deep learning image classification,which changes the feature preference of general image classification methods and makes the generation of the overall security attributes of images more scientific and reasonable.An analysis method of reordering samples according to classification probability is proposed,which intuitively explains that the feature preference of generating security attributes is closer to subjective feelings,which can effectively improve robustness.The image attribute generation experiment was conducted on Office-Caltech 10 data set.Compared with the attribute generation only using the image classification method,the accuracy of the proposed method increased by 0.66% in the training domain and 8.45% in the transfer domain,which significantly improved the robustness while improving the accuracy.2.The dynamic expansion of security attribute ensures that the security attribute database can be improved and updated continuously.Aiming at the problem that the new security attribute is difficult to be added accurately due to the static setting of image security attribute,a dynamic extension technology of image security attribute database based on dynamic group set is proposed.The dynamic adjustment of attribute structure is realized by connecting multiple image classification tasks with image retrieval task and the image security attribute database is more complete.In the stage of image retrieval database establishment,a method of image retrieval database establishment based on sample density cluster analysis is proposed,which solves the problem of image classification task lacking strict definition domain.In the phase of the combination of image classification and retrieval task,the stability condition of image classification is designed,which reduces the problem of decreasing classification accuracy caused by domain shift.According to the knowledge of image retrieval database and stability conditions,a dynamic group set with security attribute extraction elements is formed.Given a new image classification task and corresponding training data,a dynamic group database is generated through safety attribute collision detection to complete the continuous learning of image safety attributes.According to the ABAC access control model,an access control example is given.The dynamic extension experiment of multiple groups of image security attributes verifies that the technology has the ability to improve the attribute database dynamically.Compared with the single model image classification task,the accuracy of the attribute is improved by 5.56% on average.3.Security attribute recommendation is an important method for image data to obtain security attribute intelligently through security attribute data base.In view of the challenges of domain shift and semantic gap faced by traditional image attribute recommendation methods due to the wide variety of image data and the continuous emergence of a large number of unknown images under the condition of big data,an unknown image security attribute recommendation method based on attribute portrait is proposed by referring to user portrait and zero-shot learning technology.In this paper,similarity is regarded as a new attribute,and according to the different types of attributes,we design a description attribute portrait generation method based on empty field division and a measurement attribute portrait generation method based on certainty boundary,and establish a class attribute portrait.Then,on the basis of the establishment of class attribute portraits,two kinds of image security attribute recommendation methods are proposed.According to the similarity judgment results given by class attribute portraits,security attributes of unknown images are intelligently recommended.Finally,an access control implementation example based on attribute recommendation is designed.In the experiment,multiple data sets are used for unknown image attribute recommendation,and the accuracy of attribute recommendation is more than 70%,which meets the requirement of image data security attribute recommendation.4.The attribute security guarantee method is used to guarantee the credibility of the process of intelligently obtaining security attributes of image data.Aiming at the damage of adversarial example attack to image security attribute acquisition based on deep learning,an image attribute security guarantee method supporting adversarial example defense is proposed.In the stage of image preprocessing,two small size convolution kernel filtering is proposed to reduce the matching between the adversarial perturbation and the model.In the stage of image randomization,an image transformation method combining filtering and rotation with random parameters is proposed to transform the classification problem of a single image into the classification problem of multiple images after randomization,which enhances the stability of attribute extraction results.In the classification result fusion stage,the difference of multiple image features is used to improve the accuracy of attributes and realize the defense against adversarial examples.This method does not require additional model training costs and does not depend on specific data sets,which provides security for image attribute extraction.Several models and data sets are used to carry out adversarial example defense experiments.The accuracy rate of the proposed method is more than10% higher than that of similar defense methods under gray box attack,and the transferability of adversarial example is significantly reduced under white box attack,ensuring the security of attributes. |
PDF Full Text Request