Privacy-Safe Image Feature Extraction And Application

With the rise of multimedia social network,the amount of multimedia images has exploded.Some resource-constrained image owners tend to outsource their heavy image processing to the cloud.Since the images may contain some sensitive information of the owner,outsourcing the raw image data directly to untrusted cloud services may raise privacy concerns.Therefore,privacy has become an important research topic in order to achieve secured application of images.This dissertation conducts a study on image feature extraction based on Vector Homomorphic Encryption(VHE)and some key issues of its relevant application.Based existing researches,this dissertation studies key techniques such as feature extraction and recognition of encrypted images based on VHE.The main content of work and results cover the following aspects:1.A study on integer vector homomorphic encryption schemes for image processing.In this aspect,the paper focuses on a secure and efficient vector homomorphic encryption scheme,which is suitable for encryption of image matrices and feature vectors.It supports common image processing such as weighted inner product and linear transformation.Based on this,it studies the similarity comparison of ciphertext vectors,using homomorphic encryption throughout the whole image processing.2.An improved solution based on VHE for the extraction of grayscale histogram features.We propose a solution for histogram feature extraction combined with the VHE scheme and it achieve the extraction of image grayscale histogram in the ciphertext domain,with the privacy of the original image information protected.In the experiment,we then evaluate the performance of the entire algorithm and model.The results show that the extracted histogram features from the ciphertext domain can reach the same accuracy as those from the plaintext domain,with privacy of the image users protected.3.An improved solution based on VHE for HOG feature extraction.We propose a solution which is based on VHE operations supported by the vector homomorphic encryption algorithm and find a HOG algorithm that is adapted to homomorphic operations with the original HOG algorithm is simplified and improved,and then conduct an HOG feature extraction from the ciphertext domain.We also construct a pedestrian detection model with HOG feature vectors extracted respectively from the plaintext domain and ciphertext domain,using the support vector machine(SVM)algorithm.The experiments indicate that the performance of HOG algorithm in the ciphertext domain is almost equivalently effective to the original4.An application of privacy-preserving image feature extraction.We implement an image recognition experiment using grayscale histogram features and an image search experiment using HOG features previously extracted from the ciphertext domain.As what the experimental results show,in the former,homologous images are recognized,and in the latter,images similar to the targeted images are found,with the privacy of the image owners protected.