Realistic image forgeries involve a combination of splicing, resampling, cloning, region removal and other methods. While resampling detection algorithms are effective in detecting splicing and resampling, copy-move detection algorithms excel in detecting cloning and region removal. In this paper, we combine these complementary approaches in a way that boosts the overall accuracy of image manipulation detection. We use the copy-move detection method as a pre-filtering step and pass those images that are classified as untampered to a deep learning based resampling detection framework. Experimental results on various datasets including the 2017 NIST Nimble Challenge Evaluation dataset comprising nearly 10,000 pristine and tampered images shows that there is a consistent increase of 8%-10% in detection rates, when copy-move algorithm is combined with different resampling detection algorithms.

In this work, we introduce a new method for localizing image manipulations in a single digital image, such as identifying added, removed (spliced or in-painted), or deformed objects. The method utilizes the so-called Linear Pattern (LP) of digital images as a global template whose integrity can be assessed in a localized manner. The consistency of the linear pattern estimated from the image noise residual is evaluated in overlapping blocks of pixels. The manipulated region is identified by the lack of similarity in terms of the correlation coefficient computed between the power spectral density (PSD) of the LP in that region and the PSD averaged over the entire image. The method is potentially applicable to all images of sufficient resolution as long as the LP in the unmodified parts of the image has different spectral properties from that in the tampered area. No side information, such as the EXIF header or the camera model, is needed to make the method work. Experiments show the capability and limitations of the proposed method, which is robust to mild JPEG compression.

Visual content like digital images and videos are helpful in forensic investigation, which usually provides direct evidence. However, the privacy issues arising therefrom are rarely addressed. In this paper a partial encryption based scheme is proposed to enable privacy-preserving forensics for JPEG images. Viewing sensitive regions, e.g. human faces, is only granted by the trusty party when the content is proved to be of potential relevance to the investigation. A key management protocol is defined for access authorization, which ensures access to the restricted content only possible under agreement by pre-defined parties. A fully reversible partial encryption approach is applied to ensure that the encrypted regions can be perfectly recovered after the decryption is approved. Evaluation results demonstrate the applicability and effectiveness of the proposed scheme.

Convolutional neural networks offer much more accurate detection of steganography than the outgoing paradigm - classifiers trained on rich representations of images. While training a CNN is scalable with respect to the size of the training set, one cannot directly train on images that are too large due to the memory limitations of current GPUs. Most leading network architectures for steganalysis today require the input image to be a small tile with 256 × 256 or 512 × 512 pixels. Because detecting the presence of steganographic embedding changes really means detecting a very weak noise signal added to the cover image, resizing an image before presenting it to a CNN would be highly suboptimal. Applying the tile detector on disjoint segments of a larger image and fusing the results bring a plethora of new problems of how to properly fuse the outputs. In this paper, we propose a different solution to this problem based on modifying an existing leading network architecture for steganalysis in the spatial domain, the YeNet, to output statistical moments of feature maps to the fully-connected classifier part of the network. On experiments in which we adjust the payload with image size according the square root law for constant statistical detectability, we demonstrate that the proposed architecture can be trained to steganalyze images of various sizes without any or only a small loss with respect to detectors trained for a fixed image size.

The reliability of many digital forensic techniques can be negatively affected by geometrical transforms applied to the image under investigation because it breaks pixel-topixel synchronization needed for example for forensic methods that rely on sensor fingerprints. The geometrical transform typically needs to be detected and then its parameters estimated to allow subsequent successful and reliable forensic analysis. This paper focuses on blind detection of rotation and estimation of the angle by which the image was rotated. The proposed method utilizes the so-called Linear Pattern (LP) as a global template. In particular, no side information, such as watermark or the EXIF header, is required. The method is generally applicable whenever the image under investigation had a sufficiently strong LP before rotation. The performance of the method is assessed experimentally and by comparing to previous art. The main advantage of the proposed method is its accuracy for estimating small rotation angles (less than 3 degrees). It will also work after resizing.

Barcodes and watermarks offer different trade-offs for carrying data through a displayed image. Barcodes offer robust detection and decoding with high data capacity but are visually obtrusive. Watermarks are imperceptible but their detection and decoding is less robust, and they offer lower data capacity. Image-barcodes straddle this trade-off by attempting to reduce perceptual obtrusiveness compared with conventional barcodes, while minimally compromising data robustness. We propose an image-barcode for display applications that is simple, yet novel. The proposed method encodes the data into a monochrome barcode and embeds it into a suitably chosen region in the blue/red channel of a displayed image. The reduced sensitivity of the human visual system to changes in these channels (particularly the blue channel) reduces the perceptual impact of the image barcode compared with conventional black and white barcodes. The data can, however, still be robustly recovered from a typical color image capture of the displayed image-barcode by decoding only the channels with the embedding. We assess visual distortion and robustness of data recovery for the proposed method and experimentally compare against a baseline black and white barcode for two barcode modes representative of potential usage scenarios. Visual distortion for the proposed method is significantly better. Under typical settings, the proposed method introduces a mean SCIELAB-CIEDE2000 distortion of ΔE = 0.39 for the blue channel embedding and ΔE = 0.35 for the red channel embedding, compared with ΔE = 0.59 for the baseline method. For data recovery, the blue and red channel embeddings using the proposed method match the 100% decoding success rate and synchronization success rate for the baseline method, although, the pre-error-correction observed mean bit error rate of 0.047% (0.08%) for the blue (red) channel embedding is marginally worse than the performance of the baseline method.

The goal of quantitative steganalysis is to provide an estimate of the size of the embedded message once an image has been detected as containing secret data. For steganographic algorithms free of serious design flaws, such as schemes based on least significant bit replacement, the most competitive quantitative detectors have traditionally been built as regressors in rich media models. Considering the recent advances in binary steganalysis due to deep learning, in this paper we use the features extracted from the activation of such CNN detectors for the task of payload estimation. The merit of the proposed architecture is demonstrated experimentally on steganographic algorithms operating both in the spatial and JPEG domain.

Determining which processing operations were used to edit an image and the order in which they were applied is an important task in image forensics. Existing approaches to detecting single manipulations have proven effective, however, their performance may significantly deteriorate if the processing occurs in a chain of editing operations. Thus, it is very challenging to detect the processing used in an ordered chain of operations using traditional forensic approaches. First attempts to perform order of operations detection were exclusively limited to a certain number of editing operations where feature extraction and order detection are disjoint. In this paper, we propose a new data-driven approach to jointly extract editing detection features, detect multiple editing operations, and determine the order in which they were applied. We design a constrained CNN-based classifier that is able to jointly extract low-level conditional fingerprint features related to a sequence of operations as well as identify an operation's order. Through a set of experiments, we evaluated the performance of our CNN-based approach with different types of residual features commonly used in forensics. Experimental results show that our method outperforms the existing approaches.

The amount of digital imagery recorded has recently grown exponentially, and with the advancement of software, such as Photoshop or Gimp, it has become easier to manipulate images. However, most images on the internet have not been manipulated and any automated manipulation detection algorithm must carefully control the false alarm rate. In this paper we discuss a method to automatically detect local resampling using deep learning while controlling the false alarm rate using a-contrario analysis. The automated procedure consists of three primary steps. First, resampling features are calculated for image blocks. A deep learning classifier is then used to generate a heatmap that indicates if the image block has been resampled. We expect some of these blocks to be falsely identified as resampled. We use a-contrario hypothesis testing to both identify if the patterns of the manipulated blocks indicate if the image has been tampered with and to localize the manipulation. We demonstrate that this strategy is effective in indicating if an image has been manipulated and localizing the manipulations.