A Study On Privacy-preserving Outsourcing Computation Of Histogram Of Oriented Gradients In The Cloud

Large-scale multimedia data generated in our daily life has increasingly motivated companies or researchers to discover valuable knowledge and hidden information over these multimedia data by machine learning techniques.Apparently,it is inherently a time-consuming and computation-intensive task to manipulate these multimedia data locally.Thanks to the rapid development of cloud computing,companies or researchers tend to outsource these multimedia data,along with the computational intensive pro-cessing tasks,onto the cloud by leveraging its abundant resources for cost saving and flexibility.Meanwhile,the privacy issues also arise for the abuse of sensitive information contained in the outsourced multimedia data.As a preprocessing step to machine learn-ing,feature extraction has aroused new research interest on privacy-preserving compu-tations over outsourced multimedia data for its effectiveness in removing irrelevant and redundant data,increasing learning accuracy,and improving result comprehensibili-ty.In this paper,we focus on the problem of securely outsouring the famous feature extraction algorithm-Histogram of Oriented Gradient(HOG)to the untrusted cloud servers and propose two privacy-preserving HOG outsourcing solutions.The proposed solutions can well preserve the key characteristics of the extracted features as much as possible while achieving both security and effectiveness requirements.One of the proposed solutions is based on homomorphic encryption scheme,and can extract the descriptors over the encrypted image data with the help of only one cloud server,by using the properties of somewhat homomorphic encryption(SHE)in-tegrated with single-instruction multiple-data(SIMD)and tuning every step of the HOG to adapt it to the ciphertext domain.The other solution is based on secure multiparty computation(SMC)scheme.Different from the first one,there exists tow non-colluding cloud servers,and we also propose a secure comparison protocol that can compare multiple pairs of integers simultaneously with privacy preservation.The data owner randomly splits the image data,and sends to the two independent cloud servers,respectively.Then the cloud servers cooperate with each other to extract the features over the ciphertext domain,by leveraging the proposed secure comparison protocols.We give a explicit security and effectiveness analysis to each solution and conduct ex-perimental evaluations to compare our solutions with the original HOG algorithm and the state-of-the-art solutions on the real-world datasets.The results show that our so-lutions can well preserve the key characteristics of the features and perform comparably to the original HOG when being applied to human detection.