Abstract: Technologies are generally presented for employing enhanced expectation maximization (EEM) in image retrieval and authentication. Using uniform distribution as initial condition, the EEM may converge iteratively to a global optimality. If a realization of the uniform distribution is used as the initial condition, the process may also be repeatable. In some examples, a positive perturbation scheme may be used to avoid boundary overflow, often occurring with the conventional EM algorithms. To reduce computation time and resource consumption, a histogram of one dimensional Gaussian Mixture Model (GMM) with two components and wavelet decomposition of an image may be employed.

Abstract: Technologies are generally presented for employing enhanced expectation maximization (EEM) in image retrieval and authentication. Using uniform distribution as initial condition, the EEM may converge iteratively to a global optimality. If a realization of the uniform distribution is used as the initial condition, the process may also be repeatable. In some examples, a positive perturbation scheme may be used to avoid boundary overflow, often occurring with the conventional EM algorithms. To reduce computation time and resource consumption, a histogram of one dimensional Gaussian Mixture Model (GMM) with two components and wavelet decomposition of an image may be employed.

Abstract: Embodiments of the invention are directed toward methods for steganalysis that improve the detection of steganography in texture images. The methods combine features extracted from the image spatial representation and from a block discrete cosine transform (BDCT) representation with multiple different block sizes (i.e., N×N) to improve detection of data hidden in texture images. Image data that is to undergo steganalysis can be provided to embodiments of the invention in at least one of spatial (pixel) and JPEG format. When JPEG format is provided, the image is first decompressed to spatial representation, from which the features are extracted when the block size is at least one 2×2, 4×4, and 16×16. When the block size is 8×8, the JPEG coefficients (JPEG quantized 8×8 BDCT coefficients) derived directly from the JPEG image are used to extract features. In addition, the method is also effective as a universal steganalyzer for both texture and smooth/non-texture images.

Abstract: Implementations of the detection of double MPEG compression using digit based statistics are disclosed.

Abstract: Briefly, in accordance with one embodiment, a method of identifying marked content is described.

Abstract: Techniques described herein are generally related to steganalysis of suspect media. Steganalysis techniques may include receiving instances of suspect media as input for steganalytic processing. A first set of quantized blocks of data elements may be identified within the media, with this first set of blocks being eligible to be embedded with steganographic data. A second set of quantized blocks of data elements may be identified within the media, with this second set of blocks being ineligible to be embedded with steganographic data. The steganalysis techniques may requantize the first and second blocks. In turn, these techniques may compare statistics resulting from requantizing the first block with statistics resulting from requantizing the second block. The steganalysis techniques may then assess whether the first block of data elements is embedded with steganographic features based on how the statistics of the second blocks compare with the statistics of the first blocks.

Abstract: Briefly, in accordance with one embodiment, a method of identifying marked images based at least in part on frequency domain coefficient differences is disclosed.

Abstract: A method of hiding data comprising creating a histogram of an attribute of a first data set. The histogram includes occurrences of the attribute. Two adjacent occurrences are selected where a count of one of the two adjacent occurrences is zero. A second data set is embedded in data of the first data set associated with the selected adjacent occurrences.

Abstract: Techniques are generally described to determine whether a JPEG image has undergone two compressions. Probabilities can be computed for the first digits of quantized DCT (discrete cosine transform) coefficients from a set of AC (alternate current) modes to detect or determine whether the JPEG image has undergone two compressions. The set of AC modes may include a predetermined number of distinguishable AC modes where a distinguishable AC mode may be an AC mode in which a second quantization step (QS2) is not an integer multiple of the first quantization step (QS1). Classifiers may be created during a training process, and later may be used to assist in determining whether a suspect JPEG image has undergone two compressions. When the classifiers support a multi-classification system, described detection techniques may also be arranged to determine a primary quality factor for the double compressed JPEG image.

Abstract: Embodiments related to data hiding using wavelet transforms and histogram shifting are disclosed.

Abstract: A method of processing images, including: training an image classifier to obtain a trained classifier, the training including: forming multiple prediction error sets from neighboring samples of a set of known images, a prediction error for each pixel of the error sets being formed by subtracting a predicted pixel value from an original value; thresholding the formed prediction error sets; and training the image classifier using the thresholded prediction error sets.

Abstract: Techniques are generally described for detecting double JPEG compression in images. Example detection techniques of double JPEG compression may include receiving JPEG images for analysis and extracting 2-dimensional (2-D) arrays of JPEG coefficients from the images. 2-D difference arrays may be generated from the array of JPEG coefficients, with the entries in the difference array reflecting relative changes in values of pairs of entries in the array of JPEG coefficients. The detection techniques also model the difference arrays using random processes, and evaluate whether the random processes reveal statistical artifacts in the JPEG images. These statistical artifacts result from double JPEG impression performed on the JPEG images.

Abstract: The subject matter disclosed herein relates to techniques for detecting tampering of digital image data.

Abstract: Methods for reversible data hiding are disclosed herein. These methods may be applied to images for reversible image data hiding. In one embodiment, a method for reversible data hiding comprises identifying carrier data having a plurality of components and establishing prediction-errors for each of the components of the carrier data. A first threshold and a second threshold are established. The prediction-errors of the components are evaluated against the first threshold to identify components for possible hiding. The components identified for possible hiding are evaluated against the second threshold to identify components for hiding. To-be embedded data is embedded into the carrier data identified for hiding to create marked data. The marked data is evaluated for overflow or underflow. If overflow or underflow is detected, histogram modification is performed. Marked data is then established.

Abstract: Briefly, embodiments of a method of identifying marked images, in which higher order statistical moments based at least in part on a JPEG array are employed, is described.

Abstract: Computer graphics may be detected in digital images by extracting a first set of features from an input digital image, extracting a second set of features from a prediction-error image derived from the input digital image, and applying a classification algorithm to the first set of features and the second set of features to determine if the combined sets of features indicate that the input digital image corresponds to computer graphics.

Abstract: Embodiments of the invention are directed toward methods for an effective blind, passive, splicing/tampering detection. The methods of the various embodiments of the invention use a natural image model to detect image splicing/tampering with a model that is based on statistical features extracted from a given test image and multiple 2-D arrays generated by applying the block discrete cosine transform (BDCT) with several different block-sizes to the test images. Experimental results have demonstrated that the new splicing detection scheme outperforms state-of-the-art methods by a significant margin when applied to the Columbia Image Splicing Detection Evaluation Dataset.

Abstract: Briefly, in accordance with one embodiment, a method of identifying marked images based at least in part on frequency domain coefficient differences is disclosed.

Abstract: Embodiments of the invention are directed toward reversible/invertible and lossless, image data hiding that can imperceptibly hide data into digital images and can reconstruct the original image without any distortion after the hidden data have been extracted in various digital image formats including, but not limited to Joint Photographic Experts Group (JPEG). In particular, embodiments of the invention provide a lossless data hiding technique for JPEG images based on histogram pairs. that embeds data into the JPEG quantized 8×8 block DCT coefficients and achieves good performance in terms of peak signal-to-noise ratio (PSNR) versus payload through manipulating histogram pairs with optimum threshold and optimum region of the JPEG DCT coefficients. Furthermore, the invented technology is expected to be able to apply to the I-frame of Motion Picture Experts Group (MPEG) video for various applications including annotation, authentication, and forensics.

Abstract: Embodiments of the invention are directed toward methods for steganalysis that improve the detection of steganography in texture images. The methods combine features extracted from the image spatial representation and from a block discrete cosine transform (BDCT) representation with multiple different block sizes (i.e., N×N) to improve detection of data hidden in texture images. Image data that is to undergo steganalysis can be provided to embodiments of the invention in at least one of spatial (pixel) and JPEG format. When JPEG format is provided, the image is first decompressed to spatial representation, from which the features are extracted when the block size is at least one 2×2, 4×4, and 16×16. When the block size is 8×8, the JPEG coefficients (JPEG quantized 8×8 BDCT coefficients) derived directly from the JPEG image are used to extract features. In addition, the method is also effective as a universal steganalyzer for both texture and smooth/non-texture images.