Sharif Digital Repository

Availability of video recording instruments and ease of working with video editing tools have made contents of this digital signal unreliable in general. The goal of this thesis is to present a method to detect tampering of compressed videos in H.264/AVC format and restoring an approximate version of its original contents using watermarking. In the proposed scheme, a low resolution image from a number of video frames in certain time slots are embedded into the DCT coefficients of the other parts of the video which are adequately far from the reference frames. For detecting temporal and spatial tampering, the index of each frame and macroblock is embedded into itself as an authentication... 

Existing face forgery detection methods mainly focus on certain features of images, such as features related to image noise, local textures or frequency statistics of images for forgery detection. This makes the extracted representations and the final decision depend on the data in the database and makes it difficult to detect forgery with unknown manipulation methods. Solving this challenge, which is called the generalization challenge in artificial intelligence literature, has become the main goal of researchers in this field. In this thesis, the focus is on extracting effective features for success in forgery detection and preventing the performance of the forgery detection network from... 

One of the most important problems in machine learning is investigating the performance of the learning algorithms, and especially deep neural networks, on adversarial examples, which are generated by imperceptibly perturbing input images, so that cause the model make a wrong prediction. Not only this line of research is important for making deep neural networks dependable, but also can help with understanding the fundamental limitations of deep neural networks, and the nature of their operation, which can in turn provide researchers with valuable insights into artificial intelligence. In this research work, we have tried to approach the topic with a mainly theoretical mindset. The method we... 

Today, with the increasing proliferation of digital tools, manipulating digital images has become a simple matter. Therefore, in many cases related to Forensics, issues related to the intellectual property, we need to verify the authenticity of the images. There are different types of image manipulation, but image splicing manipulation is the most frequent among the types of manipulations due to its simplicity and availability.In many applications, in addition to detection, the localization of the manipulated part, which is considered segmentation at the pixel level, is also of great importance.In this project, by using a structure based on deep encoder networks, a method for improving the... 

Nowadays, methods based on deep learning networks are the most effective artificial in­ telligence methods. Although they have achieved success in various fields (such as machine vision and object recognition), practical and experimental cases show the fragility of deep learning networks against perturbations and unwanted changes of the input pattern. All these perturbations must be in a way that the main class of the perturbed input pattern can be rec­ ognized by human, but the network makes a mistake in recognizing its correct class. This thesis seeks a more accurate evaluation by designing adversarial attacks such that the main class of the adversarial pattern is detectable by human... 

Among all biometric approaches, gait analysis is one of the most practical methods for human identity recognition. Gait has a lot of advantages over other biometrics like face recognition, iris recognition, fingerprint, etc. First and foremost, the gait data can be collected from a distance, and there is no need for subject’s cooperation. Another advantage of this biometric method is its cost-effectiveness and the fact that it does not need high-resolution images. But there are significant challenges in detecting and analyzing this feature. One of the most important challenges is decreased recognition accuracy caused by identity-irrelevant factors like camera viewpoint and changes in walking... 

DeepFake is a type of forgery that uses deep learning algorithms to make changes to audio and video content that the audience is unable to detect. Nowadays, due to the threats posed by the use of DeepFake to move people's faces in video, researchers' attention has been drawn to designing methods to detect this type of forgery. Detection methods are usually classified into two types. The first case is the extraction of features to detect forgery distortions, for example, the extraction of facial orientations to detect inconsistencies. The second case is the use of deep learning networks for feature extraction and classification, of which the EfficientNet network is an example. Despite the... 

Speech is the most natural and widely used approach to communication between people and the fastest communication between humans and computers. Advances have been made in this area to achieve complete success in the natural human-computer relationship. The big challenge in this way is the inability of the computer to recognize the user's feelings; Therefore, in speech processing, one of the things that should be studied and considered is; detection of emotion from a speech by a computer. This is because Emotion recognition of speech can help extract meanings and improve the functioning of the speech recognition system. This study first defined the emotions and materials needed to build... 

Identity verification is a key part of identity reidentification process. Nowadays, Identity reidentification using face-based algorithms are popular in learning and vision area due to their generality and accessibility of this body organ. Using a fake image, occlusions on face and appearance changes like makeup can cause distortion in face verification systems which can be a drop in function of such systems. Most of these face verification models like DeepFace, FaceNet, ArcFace and SphereFace use convolution networks as their major architecture. These models, in addition to their large storage consuming and high computational complexity, due to using face as their major feature, are not... 

Talking face synthesis is a process in which is made using audio-visual data or its features. Because the face is the first output, face animation plays a crucial role in this process. A high-quality face, a balance between different facial regions, natural movements of facial organs, and the like are basic requirements to synthesize a relatively realistic talking face. There are a wide variety of applications for the photo-realistic talking face. For instance, as a teaching assistant, or reading emails and e-books are only two simple ones to mention. To reach a realistic talking face with mentioned necessary requirements, we set a goal to consider all face regions and their movements. To... 

Deep neural networks have not progresses comparative until last decade due to computational complexity and principal challenges as gradient vanishing. Thanks to newly designed hardware architecture and great breakthroughs in 2000s leading to the solution of principal challenges, we currently face a tsunami of deep architecture utilization in various machine learning applications. Sparsity of a representation as a feature to make it more descriptive has been considered in different deep learning architectures leading to different formulations where sparsity is impose on specific representations. Due to the gradient based optimization methods for training deep architecture, smooth regularizers... 

Steganalysis is the process of discovering existence of hidden data in a file or image. Recent advances in steganographic methods result in their wide use in secure communications applications. Reports of using information hiding by criminals to send messages, and malwares to communicate to servers, emphasizes importance of steganalysis in real time applications. Due to wide use of JPEG images in social media and internet, steganalysis of JPEG images is a highly active research topic in information hiding. Number of steganographic tools available for steganography in JPEG domain, shows importance of steganalysis of this image format. Classic Steganalysis techniques combine feature... 

Pitch frequency is one of the most important attributes of speech, which has been found to be quite challenging in noisy conditions. In this paper, we propose a pitch detection method based on separation of the low pitch from high pitch signals, depending on the pitch frequency below or over 200Hz, respectively, using a deep convolutional neural network. The pitch frequency is initially estimated, employing a conventional pitch detection method. From this initial estimation and using a deep convolutional neural network which determines the signals type (high-pitch or low-pitch), the pitch candidates are derived. To choose the true pitch values, we use three features in addition to soft... 

The purpose of this project is to improve current speaker verification techniques with short utterance using phonetic information extraction. I-vector technique is widely used in speaker verification systems. Different speakers span a subspace of universal acoustic space, which is usually modeled by “Universal Background model”. Speaker-specific subspace depends on the voice of speaker. In state-of-the-art speaker verification systems i-vectors are extracted by a factor analysis technique to represent speaker characteristics. Studies demonstrate that voiced phonemes contain more speaker-specific information than unvoiced. In this thesis we have classified voiced frames in order to exploit... 

Recognition of complex visual events has attracted much interest in recent years. Compared to somehow similar tasks like action recognition, event recognition is much more complex, primarily because of huge intra-class variation of events, variable video durations, lack of pre-imposed video structures, and severe preprocessing noises. To deal with these complexities and improve the state of the art approaches to the problem of video understanding, this thesis is focused on video event recognition based on frame level CNN descriptors. Using transfer learning, the image trained descriptors are applied to the video domain to make event recognition feasible in scenarios with limited... 

Nowadays, using of digital cameras is streaming across the world dramatically. Application of these devices is very diverse. One of the most interesting application of those is face verification. For example, imagine your smartphone has an application which verifies faces in front of its front camera, if that face be your face (with variation from original) then application automatically unlocks your phone. Face verification systems are also deployed in airports to verify passport photos and in smart homes. One of the most regular problems in face verification is occlusion. When your face is occluded with natural or random changes we can say your face is occluded. All of the recent papers... 

Owing to the unprecedented increasing rate of introducing new means of editing and transferring videos, appropriate and efficient methods for authentication and tampering detection are already playing a pivotal role in today's world. Amongst various methods for authentication, tampering detection and property right protection, the focus of this thesis is on watermarking. In the first place, we will present watermarking methods based upon H.264 standard. Due to the fact that almost all of the video products are stored and distributed in compressed file formats, the ability to retrieve the watermark after video compression is of crucial importance. Hence watermarking whilst video compression... 

This thesis explored and described a new steganalysis system modeling. Image steganalysis systems are divided into two parts: feature extraction from images and classification of images. Many researches have been done in both parts and satisfactory results have also been reported. There are acceptable steganalysis methods which work accurately in high rates of steganography; however steganography in low rates is still undetectable. By lowering the rate of steganography in images, the difference between stego images and clean images would be reduced which accordingly led to the reduction of the difference between the corresponding extracted features. Thus, the ability of classification... 

This project aims at forgery detection in digital videos. Most of existing methods are based on similar methods for image forgery detection. Therefore, they do not have sufficient accuracy in case of video forgery detection. In this project, we focus on local copy/move attacks in digital videos and propose 3 solutions for 3 problems in this field: 1) detection of copy/move along time axis (temporal copy/move), 2) detection of copy/move along x and y axes (spatial copy/move) and 3) detection of original and fake part in case of finding a duplication. For each of these 3 problems a feature extraction algorithm and a forgery detection algorithm are proposed. Feature extraction algorithms are... 

The rapid increase in the size of digital multimedia data collections has resulted in wide availability of multimedia contents to the general users. Effective and efficient management of these collections is an important task that has become a focus in the research of multimedia signal processing and pattern recognition. In this thesis, we address the problem of automatic classification of music, as one of the main multimedia signals. In this context, music genres are crucial descriptors that are widely used to organize the large music collections. The two main components of automatic music genre classification systems are feature extraction and classification. While features are a compact...