Information Hiding And Protection Based On Physical And Visual Perception Model

Information protection is a topic that should be endlessly studied and discussed.From ancient times to modern lives,people created many forms of methods to protect private infor-mation.As the development of modern techniques,the world has been filled with information that cannot be shared in public.During the transportation of the information,risks arise that attackers may steal private information via a variety of methods.Traditional attacks are through the Internet,where digital signals are intercepted and cracked.Conventional means to protect the information includes Cryptography and Steganography.In this thesis,we study the problem of information protection via multimedia contents through the techniques of both Cryptography and Steganography.Cryptography aims to study the encryption methods for the information.From the two aspects of Cryptography and Steganography,we concentrate on password protec-tion via modeling its attacking process,as well as new visual information hiding methods.First,we model the physical password breaking process to help protect private information.In the study,attackers are assumed to steal users' passwords by observing the trace of temperature change on the keyboard.An automated method is designed to estimate the password by analyzing local feature change of the ROI region in the thermal images.They verify that by using thermal cameras,attackers have a good chance to steal the password.We explore the feasibility of thermal attacks by analyzing the thermal sequence instead of a single image.We first analyze the process of thermal sequence signal decreasing.Then,we estimate the password by modeling with maximum likelihood algorithm.There remain several challenges to be solved.There are two challenges of the problem.The first one is to model the process,and the second one is to solve the problem with an elegant and effective solution.We study the phenomenon that the thermal camera captures the users' key touch trace and solve the two challenges above.With help of our physical model,we study the strategies to protect the password against the thermal attack.Second,we study perceptual information hiding for one single user.The algorithm is inspired by the visual masking effects.There are usually concerns that others will peep and steal private information.We make use of the multi-channel perceptual relationship of the images for private information hiding.With visual masking effect,the perceptual information can be hidden in the misleading image,while the target users can easily get the perceptual information with simply a pair of color filtering glasses.We use the visual masking effect to disturb the global and local attention of peepers,thus protecting the private information from being stolen.For the authorized user of the system,a pair of color filtering glasses is provided.Disturbing information has been filtered out by the glasses,and the remaining image mainly conveys the private information that the system wants to protect.For the peepers,however,they can hardly distinguish the protected information from the disturbing image due to visual masking effect.Third,we design an information hiding algorithm for multiple users based on TPVM technology.Temporal Psycho-visual Modulation(TPVM)[5]is recently proposed as a new dis-playing technology.With TPVM,we display distinguishable images for different users through the same displaying device.TPVM can be formulated as a nonnegative matrix factorization(NMF)problem.Traditional NMF algorithms suffer from slow convergence speed,which is not suitable for practical usage in TPVM.On the other hand,the image sources are always chang-ing.This requires the NMF algorithm to be able to generalize well in new data.Motivated by these two problems,we propose the deep neural network based model for NMF.We work on image-based NMF using deep neural network.We decompose the original signal into several bases.The base signals can be then used as features for other tasks.By combining with the traditional NMF algorithm,our method converges faster and generalize better on new data.Last,we improve the information hiding performance by analyzing and selecting the contents inside the misleading image.From experimental results,we find that the information hiding results are closely related to the contents of the misleading images.How to select the misleading images to be used in the algorithm becomes essential for the final results.The selection of the images is seen as a problem of image retrieval.By retrieving images from a large dataset,we can get candidate images that can be used as the misleading images.In the experiment,we compare several strategies of the selection of the misleading images.Finally,we combine our selection strategy with the information hiding model.