Deepfakes have gained widespread attention as they may give rise to a series of risks ranging from personal reputation damages to national security breaches. A mainstream approach to generating deepfakes is based on Generative Adversarial Networks (GAN). Various methods have been proposed to detect GAN-generated fake content. However, most of them only target a specific GAN and do not generalize well to other unseen GAN architectures. Moreover, many existing methods show poor performance on deepfakes that are imperceptible to the human eye as they heavily rely on visual artifacts, such as unblinking eyes and asymmetric faces. In this work, we exploit the spectral artifacts left by up-sampling operations that are universally used in GAN architectures for detecting high-fidelity deepfake videos. We first divide video frames into blocks containing the most informative areas, e.g., face, eyes, and mouth areas, and use their spectrum to train a ResNet-based classifier to detect GAN-generated images. Experimental results on public datasets show that our method is effective in detecting high-fidelity deepfakes and generalizes well across GANs with the same or similar up-sampling operations.

Localizing image forgeries is one of the key topics in multimedia forensics. Among many image forgery localization techniques, the one based on the photo-response non-uniformity (PRNU) noise has attracted substantial attention because of its capability of localizing forgeries regardless of the type of forgery. However, despite the devoted efforts to improving the performance of PRNU-based forgery localization, there remain challenges to be overcome, especially for detecting subtle forgeries in PRNU-attenuated regions due to complex image content. In this work, we investigate the feasibility and effectiveness of convolutional neural networks (CNN) in predicting PRNU correlations under complex backgrounds for more accurate forgery localization. The experimental results on 20 cameras and 200 realistic forgery images show that significant improvement in correlation prediction and forgery localization can be achieved even with a light-weight CNN model. The robustness of different correlation predictors against JPEG compression is also evaluated.

Most existing source-oriented image and video clustering algorithms based on sensor pattern noise (SPN) rely on the pairwise similarities, whose calculation usually dominates the overall computational time. The heavy computational burden is mainly incurred by the high dimensionality of SPN, which typically goes up to millions for delivering plausible clustering performance. This problem can be further aggravated by the uncertainty of the orientation of images or videos because the spatial correspondence between data with uncertain orientations needs to be reestablished in a brute-force search manner. In this work, we propose a rotation-invariant binary representation of SPN to address the issue of rotation and reduce the computational cost of calculating the pairwise similarities. Results on two public multimedia forensics databases have shown that the proposed approach is effective in overcoming the rotation issue and speeding up the calculation of pairwise SPN similarities for source-oriented image and video clustering.

Practical applications of digital forensics are often faced with the challenge of grouping large-scale suspicious images into a vast number of clusters, each containing images taken by the same camera. This task can be approached by resorting to the use of sensor pattern noise (SPN), which serves as the fingerprint of the camera. The challenges of large-scale image clustering come from the sheer volume of the image set and the high dimensionality of each image. The difficulties can be further aggravated when the number of classes (i.e., the number of cameras) is much higher than the average size of class (i.e., the number of images acquired by each camera). We refer to this as the NC ≫ SC problem, which is not uncommon in many practical scenarios. In this paper, we propose a novel clustering framework that is capable of addressing the NC ≫ SC problem without a training process. The proposed clustering framework was evaluated on the Dresden image database and compared with the state-of-the-art SPN-based image clustering algorithms. Experimental results show that the proposed clustering framework is much faster than the state-of-the-art algorithms while maintaining a high level of clustering quality.

In this work, we propose a method to obtain higher quality sensor pattern noise (SPN) for identifying source cameras. We believe that some components of SPN have been severely contaminated by the errors introduced by denoising filters and the quality of SPN can be improved by abandoning those components. In our proposed method, some coefficients with higher denoising errors are abandoned in the wavelet representation of SPN and the remaining wavelet coefficients are further enhanced to suppress the scene details in the SPN. These two steps aim to provide better SPN with higher signal-to-noise ratio (SNR) and therefore improve the identification performance. The experimental results on 2,000 images captured by 10 cameras (each responsible for 200 images), show that our method achieves better receiver operating characteristic (ROC) performance when compared with some state-of-the-art methods.

Although sensor pattern noise (SPN) has been proved to be an effective means to uniquely identify digital cameras, some non-unique artifacts, shared among cameras undergo the same or similar in-camera processing procedures, often give rise to false identifications. Therefore, it is desirable and necessary to suppress these unwanted artifacts so as to improve the accuracy and reliability. In this paper, we propose a novel preprocessing approach for attenuating the influence of the non-unique artifacts on the reference SPN to reduce the false identification rate. Specifically, we equalize the magnitude spectrum of the reference SPN through detecting and suppressing the peaks according to the local characteristics, aiming at removing the interfering periodic artifacts. Combined with six SPN extractions or enhancement methods, our proposed spectrum equalization algorithm is evaluated on the Dresden image database as well as our own database, and compared with the state-of-the-art preprocessing schemes. The experimental results indicate that the proposed procedure outperforms, or at least performs comparable with, the existing methods in terms of the overall receiver operating characteristic curves and kappa statistic computed from a confusion matrix, and tends to be more resistant to JPEG compression for medium and small image blocks.

In this paper, a novel forensic method of exposing cut-and-paste image forgery through detecting contrast enhancement is proposed. We reveal the inter-channel correlation introduced by color image interpolation, and show how a linear or nonlinear contrast enhancement can disturb this natural inter-channel dependency. We then construct a metric to measure these correlations, which are useful in distinguishing the original and contrast enhanced images. The effectiveness of the proposed algorithm is experimentally validated on natural color images captured by commercial cameras. Finally, its robustness against some anti-forensic algorithms is also discussed.