Research On Perceptual Hashing Algorithm Based On Generative Adversarial Network

With the rapid development of multimedia,cloud processing,and 5G technology,more and more people are entering the information age,especially due to the improvement of infrastructure in our country.Multimedia information,particularly digital images,has become pervasive in our lives,enabling people to download,upload,and disseminate image information anytime and anywhere.The frequent exchange of digital image information has raised widespread concerns about copyright protection and information authenticity.Traditional sensitive hashing techniques,commonly used for internet technologies such as lossy compression and image cropping,are no longer suitable for image forensics.In response,researchers have proposed a new digital image processing technique called perceptual hashing,which builds upon the foundation of digital watermarking and traditional sensitive hashing.Perceptual hashing technology establishes a mapping between natural images and their perceptual feature digests,ensuring that similar images have similar perceptual hash values.Due to its fast authentication speed and high storage efficiency,perceptual hashing has garnered significant attention from researchers.This paper provides a comprehensive summary of mainstream image perceptual hashing methods from both domestic and international sources.Additionally,addressing the limitations of current algorithms,this research investigates two unsupervised perceptual hashing algorithms based on Bidirectional Generative Adversarial Networks(Bi GANs).These algorithms enhance the robustness of the algorithm under content-preserving operations and address the fragility of hash code representation.The specific research contents are as follows:(1)Considering that the primary challenges in image forensics within the current internet environment involve a wide variety and high intensity of attacks,this research adopts Bi GAN as the backbone network and introduces skip connections between the encoding and generating networks.This facilitates the transmission of feature information of different dimensions from the original image to the generating network.In other words,it combines local features of different scales with global features to promote the extraction of more robust features by the encoding network during iterative adversarial training.To reduce hash storage space and ensure hash security,the extracted robust features undergo quantization and encryption processing.The algorithm’s performance is evaluated using a large-scale image database,and the results demonstrate that the skip connectionenhanced Bi GAN perceptual hashing algorithm significantly improves the robustness of multiple content-preserving operations and achieves higher detection accuracy.(2)In the perceptual hashing research of natural scene images,a discrepancy exists between robustness and discriminability.The skip connection-enhanced Bi GAN perceptual hashing algorithm,designed to handle attacks with high intensity,may mistakenly classify images with significant content differences as similar.To address this issue,this research proposes the additional integration of Mean Squared Error(MSE)loss in both the encoding and generating networks,leveraging its pixel-wise calculation to increase the weight of image detail information during network backpropagation.Experimental results on a large-scale natural image dataset demonstrate that the perceptual hashing generated by the MSE-optimized Bi GAN significantly reduces the possibility of misjudgment while maintaining perceptual robustness.