Degrees of hyper-realism already attained by present-day deepfake technology poses one of the biggest social threats of today. Deepfakes may involve multiple forms of media including audio, video and images. While most of the literature deals with threats posed by visual synthetic media, the endeavor to uncover audio deepfakes is still evolving; it demands more extensive investigation. Our attempt towards audio deepfake detection in this article, involves investigating spectral patterns present in audio spectrograms, captured appropriately by a two-attention vision transformer model (SpecViT2A), finally exploited for discriminating synthetic audios from pristine ones. We employed our model for successful identification of audio as well as multimodal deepfakes, yielding best accuracy over 99%, F1-score 0.9911, and EER as low as 3.5 on ASVSpoof 2021.

The rapid advancement of Generative AI, especially Generative Adversarial Networks (GANs), has increased the issue of fake news on Online Social Networks (OSNs) by generating deceptive face images for social media profiles. Although existing detection methods are accurate, their effectiveness decreases when images are post-processed, which is common on OSNs. In this paper, we present LPQ-Net, a model combining Local Phase Quantization (LPQ) for feature extraction with a CNN-based classifier. We explore two variants: one sets a new benchmark in detecting StyleGAN2-generated images, and the other excels in identifying images shared on Facebook, WhatsApp, and Instagram. LPQ-Net also operates with minimal parameters, outperforming state-of-the-art methods and making it ideal for resource-constraint applications. Furthermore, our solution demonstrates its effectiveness by performing exceptionally well in detecting images generated by various Diffusion models. We further show that incorporating LPQ features into fine-tuned classifiers like ResNet50, ResNet101, InceptionV3, and DenseNet121 significantly improves performance.

The rapid progress of deepfake technology has made it an increasingly significant topic in current times. While it offers valuable applications in various contexts, a major challenge lies in discerning deepfake contents. Deepfake refers to media content that is synthetically generated or altered using advanced AI techniques, many times with malicious intent to present it as genuine. This includes synthesis or manipulation of audio, video, images, and text to produce highly realistic but fabricated content. In this paper, we work towards identifying deepfake videos generated by advanced diffusion technology, by finding inconsistencies in emotional cues between real and synthetic human videos. In our experiments, we adopt Synthesia, an SOA diffusion model based synthetic video generator, vis-a-vis real videos collected from YouTube. To discern between genuine and synthetic videos, we follow a statistical approach which provides us strong cues in terms of emotional irregularities, and paves the path for detection of diffusion model generated deepfakes in the future. Alongside, we also develop a vision transformer model for fake video detection based on temporal variation in human facial expressions, which generated an accuracy of 98.11%. Our work proves that components of emotion constitute an essential trait for identifying deepfakes, strong and effective against even the most recent classes of deepfake generators.

Creating hyper-realistic synthetic images has become effortless with tremendous development in Generative Artificial Intelligence technologies. Generative Adversarial Networks (GAN) generated synthetic images, especially non-existent face images that are visually indistinguishable from real faces, pose a severe social threat by enabling misinformation dissemination, often over online social networks and through fake social profiles. In spite of successful solutions being reported in the recent literature for detecting GAN-generated synthetic images, the performance of such schemes degrades considerably with the launch of post-processing attacks. In this work, we employ gradient of an image as the key component to detect synthetic images. According to our results, gradient proves to be a considerably efficient image derivative for synthetic image detection as well as to achieve robustness against post-processing attacks. We explore two different gradient operators and design four unique deep learning-based detection networks utilizing different gradient-based feature sets. Our solution achieves state-of-the-art (SOTA) detection accuracy (above 99%) on the test set consisting of STYLEGAN2 images and outperforms SOTA solutions for detecting post-processed and compressed images.

Adaptive PRNU Denoising (APD) is one of the strongest forms of state-of-the-art counter-forensic attacks, that has largely succeeded to prevent Photo Response Non-Uniformity (PRNU) based source identification of digital images. A major way to thwart such attack would be finding a successful mapping of a counter-forensically modified image, back to its source. In other words, such attack will be rendered obsolete if accurate source identification can be made possible with the counter-forensically modified images. For the first time, we perform a feature based identification of image sources, working directly with counter-forensically modified images, without trying to recover their authentic versions. We quantize the effects of APD on forensic source identification techniques, and explore how image texture helps in source identification of APD counter-forensic images. We propose Local Binary Pattern (LBP) extraction on APD image texture layer, the key component of our work, enabling successful counter-forensic image source attribution. Our experimental results prove that the proposed methodology is highly effective to mitigate the threats posed by APD, one of the strongest counter-forensic attacks on camera model identification today.

To bring about variation in the physical and structural properties or grade of a metal, it is made to undergo specific heat treatment procedures; which can be customized to make the metal microstructure evolve desirably, to obtain specific targeted properties. Recently, computer‐based simulations of such heat treatment procedures have become popular, however, such simulations are feasible only if the digital microstructure images are available in suitable forms (optimal digital forms of the microstructure images means the distinct grains identified and the grain boundaries demarcated, i.e., segmentation of microstructure images). To this end, the authors propose a deep learning based Generative Adversarial Network (GAN) architecture for steel microstructure image segmentation. The authors’ experimental results prove the performance efficiency of the proposed GAN model, as compared to the state‐of‐the‐art. However, the proposed network architecture requires large volumes of training data, in the form of annotated ground truth segmentation masks. The current literature lacks sufficient segmented steel microstructure images for this training, to the best of their knowledge. Hence, their second contribution in this study is the development of a Convolutional Neural Network‐based framework for sufficient ground truths generation, to aid in the proposed segmentation network training.

Recently, antiforensic methods have been proposed that invalidate most of the state-of-the-art median filter digital image forensic techniques. Also, the existing counter antiforensic methods decline noticeably when evaluated on small-sized patches in JPEG compressed images. In this letter, we have developed a robust residual dense (Neural) network-based counter antiforensic median filter detection technique that exploits local dense connection and residual learning of features for improved classification of images. Experimental results demonstrate that the proposed approach achieves superior performance to state-of-the-art techniques in detecting forgeries, even in small patches, in JPEG compressed images, for both median filtered and antiforensic median filtered images.

Source camera identification (SCI) is a forensic problem of mapping an image back to its source, often in relation to cybercrime. In this digital era, this problem needs to be addressed with the utmost care as a falsely identified source might implicate an innocent person. A very practical problem in this study is the presence of unknown models in the set of cameras under question. In other words, the images under question might not have originated from any of the camera models that are accessible to the forensic analyst, but from a different inaccessible source. Under such a circumstance, the conventional source detection techniques fail to identify the correct source, and falsely map the image to one of the accessible camera models. To address this problem, here the authors propose an SCI scheme which is capable of identifying N known (accessible) as well as K unknown (inaccessible) camera models. The authors’ experimental results prove that the proposed scheme efficiently separates the known and unknown models, and helps to achieve considerably high source identification accuracy as compared to the state‐of‐the‐art.

This article describes how digital forensic techniques for source investigation and identification enable forensic analysts to map an image under question to its source device, in a completely blind way, with no a-priori information about the storage and processing. Such techniques operate based on blind image fingerprinting or machine learning based modelling using appropriate image features. Although researchers till date have succeeded to achieve extremely high accuracy, more than 99% with 10-12 candidate cameras, as far as source device prediction is concerned, the practical application of the existing techniques is still doubtful. This is due to the existence of some critical open challenges in this domain, such as exact device linking, open-set challenge, classifier overfitting and counter forensics. In this article, the authors identify those open challenges, with an insight into possible solution strategies.

Copy–move forgery is one of the most preliminary and prevalent forms of modification attack on digital images. In this form of forgery, region(s) of an image is(are) copied and pasted onto itself, and subsequently the forged image is processed appropriately to hide the effects of forgery. State‐of‐the‐art copy–move forgery detection techniques for digital images are primarily motivated toward finding duplicate regions in an image. The last decade has seen lot of research advancement in the area of digital image forensics, whereby the investigation for possible forgeries is solely based on post‐processing of images. In this study, the authors present a three‐way classification of state‐of‐the‐art digital forensic techniques, along with a complete survey of their operating principles. In addition, they analyse the schemes and evaluate and compare their performances in terms of a proposed set of parameters, which may be used as a standard benchmark for evaluating the efficiency of any general copy–move forgery detection technique for digital images. The comparison results provided by them would help a user to select the most optimal forgery detection technique, depending on the author requirements.

Majority of the existing copy‐move forgery detection algorithms operate based on the principle of image block matching. However, such detection becomes complicated when an intelligent adversary blurs the edges of forged region(s). To solve this problem, the authors present a novel approach for detection of copy‐move forgery using stationary wavelet transform (SWT) which, unlike most wavelet transforms (e.g. discrete wavelet transform), is shift invariant, and helps in finding the similarities, i.e. matches and dissimilarities, i.e. noise, between the blocks of an image, caused due to blurring. The blocks are represented by features extracted using singular value decomposition (SVD) of an image. Also, the concept of colour‐based segmentation used in this work helps to achieve blur invariance. The authors’ experimental results prove the efficiency of the proposed method in detection of copy‐move forgery involving intelligent edge blurring. Also, their experimental results prove that the performance of the proposed method in terms of detection accuracy is considerably higher compared with the state‐of‐the‐art.

Histogram‐bin‐shifting has been previously shown to be an effective method of reversibly watermarking greyscale images. For colour image reversible watermarking, histogram‐bin‐shifting technique can be extended trivially to RGB colour space. However, direct application of histogram‐bin‐shifting to the RGB colour image components, results in relatively poor performance of the watermarking algorithm. In order to improve the performance of the algorithm in terms of embedding capacity and distortion whereas preserving the inherent computational simplicity of the histogram‐bin‐shifting technique, the authors propose a technique of shifting frequency histogram bins of transformed colour components. In this study, the authors consider the YCbCr colour‐space. Experimental results on standard test images, prove that the proposed technique achieves high embedding capacity with considerably low distortion.

In general reversible watermarking algorithms, the convention is to reject the entire cover image at the receiver end if it fails authentication, since there is no way to detect the exact locations of tampering. This feature may be exploited by an adversary to bring about a form of DoS attack. Here we provide a solution to this problem in form of a tamper localization mechanism for reversible watermarking algorithms, which allows selective rejection of distorted cover image regions in case of authentication failure, thus avoiding rejection of the complete image. Additionally it minimizes the bandwidth requirement of the communication channel.