Adversarial Training (AT) impacts different architectures in distinct ways: vision models gain robustness but face reduced generalization, encoder-based models exhibit limited robustness improvements with minimal generalization loss, and recent work in latent-space adversarial training demonstrates that decoder-based models achieve improved robustness by applying AT across multiple layers. We provide the first explanation for these trends by leveraging the manifold conjecture: off-manifold adversarial examples (AEs) enhance robustness, while on-manifold AEs improve generalization. We show that vision and decoder-based models exhibit low intrinsic dimensionality in earlier layers (favoring off-manifold AEs), whereas encoder-based models do so in later layers (favoring on-manifold AEs). Exploiting this property, we introduce SMAAT, which improves the scalability of AT for encoder-based models by perturbing the layer with the lowest intrinsic dimensionality. This reduces the projected gradient descent (PGD) chain length required for AE generation, cutting GPU time by 25–33% while significantly boosting robustness. We validate SMAAT across multiple tasks, including text generation, sentiment classification, safety filtering, and retrieval augmented generation setups, demonstrating superior robustness with comparable generalization to standard training.

Sharing methods of attack and their effectiveness is a cornerstone of building robust defensive systems. Threat analysis reports, produced by various individuals and organizations, play a critical role in supporting security operations and combating emerging threats. To enhance the timeliness and automation of threat intelligence sharing, several standards have been established, with the Structured Threat Information Expression (STIX) framework emerging as one of the most widely adopted. However, generating STIX-compatible data from unstructured security text remains a largely manual, expertdriven process. To address this challenge, we introduce AZERG, a tool designed to assist security analysts in automatically generating structured STIX representations. To achieve this, we adapt general-purpose large language models for the specific task of extracting STIX-formatted threat data. To manage the complexity, the task is divided into four subtasks: entity detection (T1), entity type identification (T2), related pair detection (T3), and relationship type identification (T4). We apply task-specific fine-tuning to accurately extract relevant entities and infer their relationships in accordance with the STIX specification. To address the lack of training data, we compiled a comprehensive dataset with 4,011 entities and 2,075 relationships extracted from 141 full threat analysis reports, all annotated in alignment with the STIX standard. Our models achieved F1-scores of 84.43% for T1, 88.49% for T2, 95.47% for T3, and 84.60% for T4 in real-world scenarios. We validated their performance against a range of open- and closed-parameter models, as well as state-of-the-art methods, demonstrating improvements of 2–25% across tasks.

We introduce a novel approach for mapping attack behaviors described in threat analysis reports to entries in an adversarial techniques knowledge base. Our method leverages a multi-stage ranking architecture to efficiently rank the most related techniques based on their semantic relevance to the input text. Each ranker in our pipeline uses a distinct design for text representation. To enhance relevance modeling, we leverage pretrained language models, which we fine-tune for the technique annotation task. While generic large language models are not yet capable of fully addressing this challenge, we obtain very promising results. We achieve a recall rate improvement of +35% compared to the previous state-of-the-art results. We further create new public benchmark datasets for training and validating methods in this domain, which we release to the research community aiming to promote future research in this important direction.

We present ProvG-Searcher, a novel approach for detecting known APT behaviors within system security logs. Our approach leverages provenance graphs, a comprehensive graph representation of event logs, to capture and depict data provenance relations by mapping system entities as nodes and their interactions as edges. We formulate the task of searching provenance graphs as a subgraph matching problem and employ a graph representation learning method. The central component of our search methodology involves embedding of subgraphs in a vector space where subgraph relationships can be directly evaluated. We achieve this through the use of order embeddings that simplify subgraph matching to straightforward comparisons between a query and precomputed subgraph representations. To address challenges posed by the size and complexity of provenance graphs, we propose a graph partitioning scheme and a behavior-preserving graph reduction method. Overall, our technique offers significant computational efficiency, allowing most of the search computation to be performed offline while incorporating a lightweight comparison step during query execution. Experimental results on standard datasets demonstrate that ProvG-Searcher achieves superior performance, with an accuracy exceeding 99% in detecting query behaviors and a false positive rate of approximately 0.02%, outperforming other approaches.

Abstract Adversarial training has been empirically shown to be more prone to overfitting than standard training. The exact underlying reasons are still not fully understood. In this paper, we identify one cause of overfitting related to current practices of generating adversarial examples from misclassified samples. We show that, following current practice, adversarial examples from misclassified samples results in harder-to-classify samples than the original ones. This leads to a complex adjustment of the decision boundary during training and hence overfitting. To mitigate this issue, we propose A3T, an accuracy aware AT method that generate adversarial example differently for misclassified and correctly classified samples. We show that our approach achieves better generalization while maintaining comparable robustness to state-of-the-art AT methods on a wide range of computer vision, natural language processing, and tabular tasks.

We study the problem of profiling news media on the Web with respect to their factuality of reporting and bias.This is an important but under-studied problem related to disinformation and "fake news" detection, but it addresses the issue at a coarser granularity compared to looking at an individual article or an individual claim.This is useful as it allows to profile entire media outlets in advance.Unlike previous work, which has focused primarily on text (e.g., on the articles published by the target website, or on the textual description in their social media profiles or in Wikipedia), here we focus on modeling the similarity between media outlets based on the overlap of their audience.This is motivated by homophily considerations, i.e., the tendency of people to have connections to people with similar interests, which we extend to media, hypothesizing that similar types of media would be read by similar kinds of users.In particular, we propose GREENER (GRaph nEural nEtwork for News mEdia pRofiling), a model that builds a graph of inter-media connections based on their audience overlap, and then uses graph neural networks to represent each medium.We find that such representations are quite useful for predicting the factuality and the bias of news media outlets, yielding improvements over state-ofthe-art results reported on two datasets.When augmented with conventionally used representations obtained from news articles, Twitter, YouTube, Facebook, and Wikipedia, we improve over previous work by 2.5-27 macro-F1 points absolute for the two tasks and datasets.

We introduce the use of video coding settings for source identification and propose a new approach that incorporates encoding and encapsulation aspects of a video. To this end, a joint representation of the overall file metadata is developed and used in conjunction with a two-level hierarchical classification method. At the first level, our method groups videos into metaclasses considering several abstractions that represent high-level structural properties of file metadata. This is followed by a more nuanced classification of classes that comprise each metaclass. The method is evaluated on more than 20K videos obtained by combining four public video datasets. Tests show that a balanced accuracy of 91% is achieved in correctly identifying the class of a video among 119 video classes. This corresponds to an improvement of 6.5% over the conventional approach based on video file encapsulation characteristics. Analysis performed on a large, unlabeled video set also confirmed the aptness of our approach. To further demonstrate the versatility of encoding parameters, we consider attribution of partial video files where file metadata is not available. Our results show that, even in this limited setting that is intrinsic to forensic file recovery, an identification accuracy of 57% can be achieved through the use of a subset of encoding parameters estimated from coded video data.

Image stabilization performed during imaging and/or post-processing poses one of the most significant challenges to photo-response non-uniformity based source camera attribution from videos. When performed digitally, stabilization involves cropping, warping, and inpainting of video frames to eliminate unwanted camera motion. Hence, successful attribution requires inversion of these transformations in a blind manner. To address this challenge, we introduce a source camera verification method for videos that takes into account spatially variant nature of stabilization transformations and assumes a larger degree of freedom in their search. Our method identifies transformations at a sub-frame level, incorporates a number of constraints to validate their correctness, and offers computational flexibility in the search for the correct transformation. The method also adopts a holistic approach in countering disruptive effects of other video generation steps, such as video coding and downsizing, for more reliable attribution. Tests performed on one public and two custom datasets show that the proposed method is able to verify the source of 23-30% of all videos that underwent stronger stabilization, depending on computation load, without a significant impact on false attribution.

We address the problem of decoding video file fragments when the necessary encoding parameters are missing. With this objective, we propose a method that automatically generates H.264 video headers containing these parameters and extracts coded pictures in the partially available compressed video data. To accomplish this, we examined a very large corpus of videos to learn patterns of encoding settings commonly used by encoders and created a parameter dictionary. Further, to facilitate a more efficient search our method identifies characteristics of a coded bitstream to discriminate the entropy coding mode. It also utilizes the application logs created by the decoder to identify correct parameter values. Evaluation of the effectiveness of the proposed method on more than 55K videos with diverse provenance shows that it can generate valid headers on average in 11.3 decoding trials per video. This result represents an improvement by more than a factor of 10 over the conventional approach of video header stitching to recover video file fragments.

We study the individuality of the human voice with respect to a widely used feature representation of speech utterances, namely, the i-vector model. As a first step toward this goal, we compare and contrast uniqueness measures proposed for different biometric modalities. Then, we introduce a new uniqueness measure that evaluates the entropy of i-vectors while taking into account speaker level variations. Our measure operates in the discrete feature space and relies on accurate estimation of the distribution of i-vectors. Therefore, i-vectors are quantized while ensuring that both the quantized and original representations yield similar speaker verification performance. Uniqueness estimates are obtained from two newly generated datasets and the public VoxCeleb dataset. The first custom dataset contains more than one and a half million speech samples of 20,741 speakers obtained from TEDx Talks videos. The second one includes over twenty one thousand speech samples from 1,595 actors that are extracted from movie dialogues. Using this data, we analyzed how several factors, such as the number of speakers, number of samples per speaker, sample durations, and diversity of utterances affect uniqueness estimates. Most notably, we determine that the discretization of i-vectors does not cause a reduction in speaker recognition performance. Our results show that the degree of distinctiveness offered by i-vector-based representation may reach 43–70 bits considering 5-second long speech samples; however, under less constrained variations in speech, uniqueness estimates are found to reduce by around 30 bits. We also find that doubling the sample duration increases the distinctiveness of the i-vector representation by around 20 bits.

The photo-response non-uniformity (PRNU) is a distinctive image sensor characteristic, and an imaging device inadvertently introduces its sensor's PRNU into all media it captures. Therefore, the PRNU can be regarded as a camera fingerprint and used for source attribution. The imaging pipeline in a camera, however, involves various processing steps that are detrimental to PRNU estimation. In the context of photographic images, these challenges are successfully addressed and the method for estimating a sensor's PRNU pattern is well established. However, various additional challenges related to generation of videos remain largely untackled. With this perspective, this work introduces methods to mitigate disruptive effects of widely deployed H.264 and H.265 video compression standards on PRNU estimation. Our approach involves an intervention in the decoding process to eliminate a filtering procedure applied at the decoder to reduce blockiness. It also utilizes decoding parameters to develop a weighting scheme and adjust the contribution of video frames at the macroblock level to PRNU estimation process. Results obtained on videos captured by 28 cameras show that our approach increases the PRNU matching metric up to more than five times over the conventional estimation method tailored for photos. Tests on a public dataset also verify that the proposed method improves the attribution performance by increasing the accuracy and allowing the use of smaller length videos to perform attribution.

Orphaned file fragment carving is concerned with recovering contents of encoded data in the absence of any coding metadata. Constructing an orphaned file carver requires addressing three challenges: a specialized decoder to interpret partial file data; the ability to discriminate a specific type of encoded data from all other types of data; and comprehensive prior knowledge on possible encoding settings. In this work, we build on the ability to render a partial image contained within a segment of JPEG coded data to introduce a new carving tool that addresses all these challenges. Towards this goal, we first propose a new method that discriminates JPEG file data from among 993 file data types with 97.7% accuracy. We also introduce a method for robustly delimiting entropy coded data segments of JPEG files. This in turn allows us to identify partial JPEG file headers with zero false rejection and 0.1% of false alarm rate. Secondly, we examine a very diverse image set comprising more than 7 million images. This ensures comprehensive coverage of coding parameters used by 3,269 camera models and a wide variety of image editing tools. Further, we assess the potential impact of the developed tool on practice in terms of the amount of new evidence that it can recover. Recovery results on a set of used SD cards purchased online show that our carver is able to recover 24% more image data as compared to existing file carving tools. Evaluations performed on a standard dataset also show that JpgScraper improves the state-of-the-art significantly in carving JPEG file data.

Given a query image or video, or a known camera fingerprint, there is a lack of capabilities for fast identification of media, from a large repository of images and videos, that match the query fingerprint. This work introduces a new approach that improves the computation efficiency of pairwise camera fingerprint matching and incorporates group testing to make the search more effective. More specifically, we jointly leverage the individual strengths of composite fingerprints and fingerprint digests in a novel manner and design two methods that are superior to existing approaches. The results show that under very high-performance requirements, where the probability of correct identification is close to one with a false-positive rate of zero, the proposed search methods are 2-8 times faster than the state-of-art search methods.

The photo response non-uniformity noise associated with an imaging sensor has been shown to be a unique and persistent identifier that can be treated as the sensor's digital fingerprint. The method for attributing an image to a particular camera, however, is not suitable for source identification due to efficiency considerations, which is a one-to-many matching of a single fingerprint against a database of fingerprints. To address this problem, we propose a group-testing approach based on the notion of composite fingerprints (CFs), generated by combining many actual fingerprints together into a single fingerprint. Our technique organizes a database of fingerprints into an unordered binary search tree, wherein each internal node is represented by a fingerprint composited from all the fingerprints at the leaf nodes in the subtree beneath that node. Different search strategies are considered, and the performance is analyzed analytically and verified using numerical simulations as well as experimental results. Our results are presented in comparison with the linear search-based approach that utilizes fingerprint digests for more effective computation. Results obtained under the best achievable accuracy showed that the proposed method yields a lower overall computational cost. It is also shown that by complementary use of the fingerprint dimension reduction and CF-based search tree approaches, it is possible to further improve the search efficiency.

File carving techniques allow for recovery of files from storage devices in the absence of any file system metadata. When data are encoded and compressed, the current paradigm of carving requires the knowledge of the compression and encoding settings to succeed. In this paper, we advance the state of the art in JPEG file carving by introducing the ability to recover fragments of a JPEG file when the associated file header is missing. To realize this, we examined JPEG file headers of a large number of images collected from Flickr photo sharing site to identify their structural characteristics. Our carving approach utilizes this information in a new technique that performs two tasks. First, it decompresses the incomplete file data to obtain a spatial domain representation. Second, it determines the spatial domain parameters to produce a perceptually meaningful image. Recovery results on a variety of JPEG file fragments show that given the knowledge of Huffman code tables, our technique can very reliably identify the remaining decoder settings for all fragments of size 4 KiB or above. Although errors due to detection of image width, placement of image blocks, and color and brightness adjustments can occur, these errors reduce significantly when fragment sizes are >32 KiB.

This paper presents a child pornographic image detection system that identifies human skin tones in digital images, extracts features to detect explicit images and performs facial image based age classification. The novelty of the technique relies on the use of a robust and very fast skin color filter and a new set of facial features for improved identification of child faces. Tests on a dataset containing explicit images taken under different illuminations and reflecting a diversity of human skin tones, show that explicit images can be differentiated from benign images with around 90% accuracy. Similarly, tests performed on adult and child facial images yielded an accuracy of 80% in detecting child faces. Test conducted on 105 images involving semi-naked children (with no sexual context) revealed that the system has true positive rates of 83% in detecting explicit-like images and 96.5% in detecting child faces.

We present a semi-automatic method to reconstruct shredded documents. The novelty of the method lies in the way it performs pairwise matching of chads. The technique divides chad contours into curves using corner detection and introduces a procedure to assess the match of two curves. The proposed curve matching technique is robust to translation and rotation and can cope with shape deformations due to shredding by allowing overlapping of chads during matching. The alignment of text lines, crossing characters and color information on the chads is also utilized to improve matching performance. Visual interfaces are designed to allow for user input in identifying correctly matching chad pairs and reconstructing the document. The effectiveness of the method is demonstrated by solving the first and second puzzles of the DARPA shredder challenge.

The availability of sophisticated source attribution techniques raises new concerns about privacy and anonymity of photographers, activists, and human right defenders who need to stay anonymous while spreading their images and videos. Recently, the use of seam-carving, a content-aware resizing method, has been proposed to anonymize the source camera of images against the well-known photoresponse nonuniformity (PRNU)-based source attribution technique. In this paper, we provide an analysis of the seam-carving-based source camera anonymization method by determining the limits of its performance introducing two adversarial models. Our analysis shows that the effectiveness of the deanonymization attacks depend on various factors that include the parameters of the seam-carving method, strength of the PRNU noise pattern of the camera, and an adversary's ability to identify uncarved image blocks in a seam-carved image. Our results show that, for the general case, there should not be many uncarved blocks larger than the size of $50\times 50$ pixels for successful anonymization of the source camera.

As image source attribution techniques have become significantly sophisticated and are now becoming commonplace, there is a growing need for capabilities to anonymize images and videos. Focusing on the photo response non-uniformity noise pattern based sensor fingerprinting technique, this work evaluates the effectiveness of well-established seam carving method to defend against sensor fingerprint matching. We consider ways in which seam-carving based anonymization can be countered and propose enhancements over conventional seam carving method. Our results show that applying geometrical distortion in addition to seam carving will make counter attack very ineffective both in terms of computational complexity and accuracy.

It is now established that photo-response nonuniformity noise pattern can be reliably used as a fingerprint to identify an image sensor. The large size and random nature of sensor fingerprints, however, make them inconvenient to store. Further, associated fingerprint matching method can be computationally expensive, especially for applications that involve large-scale databases. To address these limitations, we propose to represent sensor fingerprints in binary-quantized form. It is shown through both analytical study and simulations that the reduction in matching accuracy due to quantization is insignificant as compared to conventional approaches. Experiments on actual sensor fingerprint data are conducted to confirm that only a slight increase occurred in the probability of error and to demonstrate the computational efficacy of the approach.

Most work on steganalysis, except a few exceptions, have primarily focused on providing features with high discrimination power without giving due consideration to issues concerning practical deployment of steganalysis methods. In this work, we focus on machine learning aspect of steganalyzer design and utilize a hierarchical ensemble of classifiers based approach to tackle two main issues. Firstly, proposed approach provides a workable and systematic procedure to incorporate several steganalyzers together in a composite steganalyzer to improve detection performance in a scalable and cost-effective manner. Secondly, since the approach can be readily extended to multi-class classification it can also be used to infer the steganographic technique deployed in generation of a stego-object. We provide results to demonstrate the potential of the proposed approach.

Being able to measure the actual information content of biometrics is very important but also a challenging problem. Main difficulty here is not only related to the selected feature representation of the biometric data, but also related to the matching algorithm employed in biometric systems. In this paper, we propose a new measure for measuring biometric information using relative entropy between intra-user and inter-user distance distributions. As an example, we evaluated the proposed measure on a face image dataset.

Copy-move forgery is a specific type of image tampering, where a part of the image is copied and pasted on another part of the same image. In this paper, we propose a new approach for detecting copy-move forgery in digital images, which is considerably more robust to lossy compression, scaling and rotation type of manipulations. Also, to improve the computational complexity in detecting the duplicated image regions, we propose to use the notion of counting bloom filters as an alternative to lexicographic sorting, which is a common component of most of the proposed copy-move forgery detection schemes. Our experimental results show that the proposed features can detect duplicated region in the images very accurately, even when the copied region was undergone severe image manipulations. In addition, it is observed that use of counting bloom filters offers a considerable improvement in time efficiency at the expense of a slight reduction in the robustness.

In this paper, a novel individual source scanner identification scheme is proposed. The scheme uses traces of dust, dirt, and scratches over scanner platen on scanned images to characterize a source scanner. The efficacy of the proposed scheme is substantiated with experimental analysis. The robustness of the scheme to the JPEG compression is also investigated. Experimental results show that proposed scheme could be used to match a scanned image to its source.

Digital single lens reflex cameras suffer from a well-known sensor dust problem due to interchangeable lenses that they deploy. The dust particles that settle in front of the imaging sensor create a persistent pattern in all captured images. In this paper, we propose a novel source camera identification method based on detection and matching of these dust-spot characteristics. Dust spots in the image are detected based on a (Gaussian) intensity loss model and shape properties. To prevent false detections, lens parameter-dependent characteristics of dust spots are also taken into consideration. Experimental results show that the proposed detection scheme can be used in identification of the source digital single lens reflex camera at low false positive rates, even under heavy compression and downsampling.

Discrimination of computer generated images from real images is becoming more and more important. In this paper, we propose the use of new features to distinguish computer generated images from real images. The proposed features are based on the differences in the acquisition process of images. More specifically, traces of demosaicking and chromatic aberration are used to differentiate computer generated images from digital camera images. It is observed that the former features perform very well on high quality images, whereas the latter features perform consistently across a wide range of compression values. The experimental results show that proposed features are capable of improving the accuracy of the state-of-the-art techniques.

Powerful digital media editing tools make producing good quality forgeries very easy for almost anyone. Therefore, proving the authenticity and integrity of digital media becomes increasingly important. In this work, we propose a simple method to detect image tampering operations that involve sharpness/blurriness adjustment. Our approach is based on the assumption that if a digital image undergoes a copy-paste type of forgery, average sharpness/blurriness value of the forged region is expected to be different as compared to the non-tampered parts of the image. The method of estimating sharpness/blurriness value of an image is based on the regularity properties of wavelet transform coefficients which involves measuring the decay of wavelet transform coefficients across scales. Our preliminary results show that the estimated sharpness/blurriness scores can be used to identify tampered areas of the image.

In a novel method for identifying the source camera of a digital image is proposed. The method is based on first extracting imaging sensor's pattern noise from many images and later verifying its presence in a given image through a correlative procedure. In this paper, we investigate the performance of this method in a more realistic setting and provide results concerning its detection performance. To improve the applicability of the method as a forensic tool, we propose an enhancement over it by also verifying that class properties of the image in question are in agreement with those of the camera. For this purpose, we identify and compare characteristics due to demosaicing operation. Our results show that the enhanced method offers a significant improvement in the performance.

This paper focuses on a problem that is common to most watermarking-based ownership dispute resolutions and ownership assertion systems. Such systems are vulnerable to a simple but effective class of attacks that exploit the high false-positive rate of the watermarking techniques to cast doubt on the reliability of a resulting decision. To mitigate this vulnerability, we propose embedding multiple watermarks, as opposed to embedding a single watermark, and detecting a randomly selected subset of them while constraining the embedding distortion. The crux of the scheme lies in both watermark generation, which deploys a family of one-way functions and selective detection, which injects uncertainty into the detection process. The potential of this approach in reducing the false-positive probability is analyzed under various operating conditions and compared to single watermark embedding. The multiple watermark embedding and selective detection technique is incorporated analytically into the additive watermarking technique and results obtained through numerical solutions are presented to illustrate its effectiveness.

We describe a digital image forensics technique to distinguish images captured by a digital camera from computer generated images. Our approach is based on the fact that image acquisition in a digital camera is fundamentally different from the generative algorithms deployed by computer generated imagery. This difference is captured in terms of the properties of the residual image (pattern noise in case of digital camera images) extracted by a wavelet based denoising filter. In (Jan Lukas, et al., 2005), it is established that each digital camera has a unique pattern noise associated with itself. In addition, our results indicate that the two type of residuals obtained from different digital camera images and computer generated images exhibit some common characteristics that is not present in the other type of images. This can be attributed to fundamental differences in the image generation processes that yield the two types of images. Our results are based on images generated by the Maya and 3D Studio Max software, and various digital camera images.

The primary goal of image steganography techniques has been to maximize embedding rate while minimizing the detectability of the resulting stego images against steganalysis techniques. However, one particular advantage of steganography, as opposed to other information hiding techniques, is that the embedder has the freedom to choose a cover image that result in the least detectable stego image. This resource has largely remained unexploited in the proposed embedding techniques. In this paper, we study the problem of cover selection by investigating three scenarios in which the embedder has either no knowledge, partial knowledge, or full knowledge of the steganalysis technique. For example, we illustrate through experiments how simple statistical measures could help embedder minimize detectability, at times by 65%, in the partial knowledge case.

Watermarking technologies have been envisioned as a potential means for establishing ownership on digital media objects. However, achievable robustness and false-positive rates of the state-of-the-art watermarking techniques raise doubts about applicability of watermarking to ownership problem. With this perspective, we address the security weaknesses common to most watermarking techniques and assess the role of watermarking in construction of ownership assertion systems. We identify the requirements of a watermarking based ownership assertion system. Also, we provide a basic functional outline of a practical version of such a system and identify its potential vulnerabilities. To mitigate these vulnerabilities, we aim at reducing the false positive rate of the watermark detection scheme. For this purpose, we propose embedding multiple watermarks as opposed to single watermark embedding while constraining the embedding distortion. The crux of the proposed method lies in watermark generation which deploys a family of one-way functions. We incorporate the multiple watermark embedding idea with the additive watermarking technique [1] and present results to illustrate the potential of this approach in reducing the false-positive rate of the watermark detection scheme.

In this paper, we propose a novel approach to image steganography in which embedding is done without making explicit modifications to the image; that is, the embedding distortion introduced to the cover image is both perceptually and statistically ensured to be less detectable. If an image is divided into blocks and each block is hashed, then the hash values could represent the embedded message content. A set of replacement blocks can be obtained by consecutively capturing images of the same scene (or resampling the incident light). Since image content remains the same and noise is sampled, the set of replacements are statistically compatible while still providing unique hash values. With such an approach, the embedder can choose the block with hash value corresponding to the message content without violating any of the natural image statistics. When applied to JPEG images, experiments show that this technique can achieve embedding rates of 0.063 bits per DCT coefficient or 0.157 bits per nonzero DCT coefficient while still remaining undetectable by Farid's universal steganalysis tool.

In this work, we focus our interest on blind source camera identification problem by extending our results in the direction of M. Kharrazi et al. (2004). The interpolation in the color surface of an image due to the use of a color filter array (CFA) forms the basis of the paper. We propose to identify the source camera of an image based on traces of the proprietary interpolation algorithm deployed by a digital camera. For this purpose, a set of image characteristics are defined and then used in conjunction with a support vector machine based multi-class classifier to determine the originating digital camera. We also provide initial results on identifying source among two and three digital cameras.

We analyze quantization based embedding and detection schemes in terms of the data hiding framework we have introduced (Ref.1, 2). This framework enables a better connection between analytical results and practical designs. We lay out the performance evaluation criteria and present comparison results for scalar quantization based data hiding techniques. Embedder-detector designs are evaluated based on three key issues. These are: (1) the type of post-processing employed at the embedder; (2) the form of demodulation; (3) the criteria used to optimize embedding-detection parameters. Data hiding methods are compared based on rate, correlation, and probability of error performance merits.

An interesting problem in digital forensics is that given a digital image, would it be possible to identify the camera model which was used to obtain the image. In this paper we look at a simplified version of this problem by trying to distinguish between images captured by a limited number of camera models. We propose a number of features which could be used by a classifier to identify the source camera of an image in a blind manner. We also provide experimental results and show reasonable accuracy in distinguishing images from the two and five different camera models using the proposed features.

In this paper, we analyze oblivious (blind) information hiding methods from a new perspective. In 1983, Costa introduced a communications framework that also applies to oblivious information hiding. We present an alternate and equivalent framework by carrying out the channel dependent nature of the optimal encoder in a different manner. Within the proposed framework, decoder structure is simplified, although in effect it's a slender advantage compared to overall complexity. This interpretation provides a better connection between the analytical results and practical designs. We evaluate the practical embedding-detection schemes employing scalar quantization procedures along with thresholding and distortion compensation types of processings from this perspective. Furthermore, we justify the assumptions for the optimality of the two types of processings.

Compression is the most common application that an forms of multimedia data undergo. Lossy nature of the compression due to the use of quantizers should be taken into account by the watermarking technique employed. In this paper, we present a data hiding scheme that incorporates the embedding with the quantization of lossy compression. We show that embedder-detector sets making use of compression scheme's quantization characteristics have better payload and lower compression bit rates. We also optimized the system parameters that maximize the hiding rate at different compression levels.

This paper presents a simple and effective preprocessing method developed for editing compressed video sequences. The proposed method involves extracting information about different video segments from the compressed bitstream. The algorithm is not designed to distinguish among types of segments but rather to indicate the position and duration. Since no decoding of the bitstream is done, the computational load of the algorithm is very low. Although the experimental results are shown on MPEG compressed video sequences, the algorithm can easily be applied to MJPEG, MPEG4 sequences given the header information.