Spatial Domain Steganography Of High Dynamic Range Image

The development of information technology has made communication more convenient and frequent,but it has also posed significant challenges to information security.Since the development of the Internet,various information security incidents have emerged one after another,which has stimulated the development of information security technologies.Currently,in addition to the most commonly used encryption,steganography is also gradually being widely applied in information security.Unlike encryption,the primary goal of steganography is to make potential attackers unaware of the presence of secret information in the carrier,rather than making it impossible for attackers to crack the information after they become aware of its existence.With the development of the Internet and the advent of the mobile Internet era,images,as a kind of information carrier,have become increasingly important and have also become the most important carrier in steganography.High dynamic range(HDR)technology is an image technology that can present realworld scenes more realistically and has been increasingly applied in industries such as film,gaming,and advertising.Although steganography techniques for low dynamic range(LDR)images have matured after more than 20 years of development,steganography techniques for HDR images are still rare.Therefore,this thesis conducts research on spatial domain steganography techniques for HDR images and proposes new steganography methods from two perspectives: traditional algorithms and deep learning methods.(1)A dynamic embedding algorithm is proposed,which can embed secret information into HDR images in RGBE format without producing any stream expansion.This is also the first known algorithm that completely eliminates stream expansion.The dynamic embedding algorithm associates the embedding location with the secret information itself,which improves the security of steganography.In addition,this thesis also proposes a pre-processing method for carrier images,which can reduce the capacity loss caused by eliminating stream expansion from over 80% to below 1%.The algorithm uses dual keys and provides a large key space,which also provides high key security for the algorithm.(2)A deep learning steganography scheme with a hybrid channel is proposed,which utilizes the characteristics of RGBE format images,normalizes their four-channel data for training,and designs a new loss function that concatenates the three-channel and fourchannel data of the generated stego image,achieving high stego image quality.The scheme is not only difficult to distinguish visually but also superior to other similar methods in terms of comprehensive evaluation indicators.This scheme also uses an adaptive loss balance strategy to replace the fixed-weight strategy used in other research,improving the quality of the recovered secret image.Finally,the scheme uses an improved VRCNN(Variable-filter-size Residue-learning CNN)as the extraction network,greatly reducing the number of parameters,which is less than 1/3 of the fully convolutional network scheme used in similar works.