In this paper, we propose dictionary attacks against speaker verification-a novel attack vector that aims to match a large fraction of speaker population by chance. We introduce a generic formulation of the attack that can be used with various speech representations and threat models. The attacker uses adversarial optimization to maximize raw similarity of speaker embeddings between a seed speech sample and a proxy population. The resulting master voice successfully matches a non-trivial fraction of people in an unknown population. Adversarial waveforms obtained with our approach can match on average 69% of females and 38% of males enrolled in the target system at a strict decision threshold calibrated to yield false alarm rate of 1%. By using the attack with a black-box voice cloning system, we obtain master voices that are effective in the most challenging conditions and transferable between speaker encoders. We also show that, combined with multiple attempts, this attack opens even more to serious issues on the security of these systems.

Analysis of imaging sensors is one of the most reliable photo forensic techniques, but it is increasingly challenged by complex image processing in modern cameras. The underlying photo response non-uniformity (PRNU) is distilled into a static sensor fingerprint unique for each device. This makes it easy to estimate and spoof and limits its reliability in face of sophisticated attackers. We propose to exploit computational capabilities of emerging intelligent vision sensors to design next-generation computational sensor fingerprints. Such sensors allow for running neural network inference directly on raw pixels, which enables end-to-end optimization of the entire photo acquisition and distribution pipeline. Control over fingerprint generation allows for adaptation to various requirements and threat models. In this study we provide a detailed assessment of security properties and evaluate two approaches to prevent spoofing: fingerprint generation based on local image content and adversarial training. We found that adversarial training is currently impractical, but content fingerprints deliver good performance in the considered cross-domain (RAW-RGB) setting and could provide robust best-effort protection against photo manipulation. Moreover, computational fingerprints can alleviate other limitations of PRNU, e.g., its limited reliability for dark/texture content and expensive fingerprint storage that hinders scalability. To enable this line of work, we developed a novel open-source and high-fidelity simulation environment for modeling photo acquisition and distribution pipelines (https://github.com/pkorus/neural-imaging).

Can we leverage high-resolution information without the unsustainable quadratic complexity to input scale? We propose Traversal Network (TNet), a novel multi-scale hard-attention architecture, which traverses image scale-space in a top-down fashion, visiting only the most informative image regions along the way. TNet offers an adjustable trade-off between accuracy and complexity, by changing the number of attended image locations. We compare our model against hard-attention baselines on ImageNet, achieving higher accuracy with less resources (FLOPs, processing time and memory). We further test our model on fMoW dataset, where we process satellite images of size up to $896 \times 896$ px, getting up to $2.5$x faster processing compared to baselines operating on the same resolution, while achieving higher accuracy as well. TNet is modular, meaning that most classification models could be adopted as its backbone for feature extraction, making the reported performance gains orthogonal to benefits offered by existing optimized deep models. Finally, hard-attention guarantees a degree of interpretability to our model's predictions, without any extra cost beyond inference.

We present FiFTy, a modern file-type identification tool for memory forensics and data carving. In contrast to previous approaches based on hand-crafted features, we design a compact neural network architecture, which uses a trainable embedding space. Our approach dispenses with the explicit feature extraction which has been a bottleneck in legacy systems. We evaluate the proposed method on a novel dataset with 75 filetypes - the most diverse and balanced dataset reported to date. FiFTy consistently outperforms all baselines in terms of speed, accuracy and individual misclassification rates. We achieved an average accuracy of 77.5% with processing speed of ≈38 sec/GB, which is better and more than an order of magnitude faster than the previous state-of-the-art tool - Sceadan (69% at 9 min/GB). Our tool and the corresponding dataset is open-source.

Forensic analysis of digital photo provenance relies on intrinsic traces left in the photograph at the time of its acquisition. Such analysis becomes unreliable after heavy post-processing, such as down-sampling and re-compression applied upon distribution in the Web. This paper explores end-to-end optimization of the entire image acquisition and distribution workflow to facilitate reliable forensic analysis at the end of the distribution channel. We demonstrate that neural imaging pipelines can be trained to replace the internals of digital cameras, and jointly optimized for high-fidelity photo development and reliable provenance analysis. In our experiments, the proposed approach increased image manipulation detection accuracy from 45% to over 90%. The findings encourage further research towards building more reliable imaging pipelines with explicit provenance-guaranteeing properties.

In this paper, we assess vulnerability of speaker verification systems to dictionary attacks. We seek master voices, i.e., adversarial utterances optimized to match against a large number of users by pure chance. First, we perform menagerie analysis to identify utterances which intrinsically hold this property. Then, we propose an adversarial optimization approach for generating master voices synthetically. Our experiments show that, even in the most secure configuration, on average, a master voice can match approx. 20% of females and 10% of males without any knowledge about the population. We demonstrate that dictionary attacks should be considered as a feasible threat model for sensitive and high-stakes deployments of speaker verification.

In this paper, we address the problem of forensic photo carving, which serves as one of the key sources of digital evidence in modern law enforcement. We propose efficient algorithms for assembling meaningful photographs from orphaned photofragments, carved without access to file headers, meta-data, or compression settings. The addressed problem raises a novel variant of a jigsaw puzzle with an unknown number of mixed images, missing pieces, and severe brightness and colorization artifacts. We construct an efficient compatibility metric for matching puzzle pieces and a corresponding image stitching procedure which allows us to mitigate these artifacts. To facilitate photo assembly, we perform a forensic analysis of the fragments to provide clues about their location within the frame of the imaging sensor. The proposed algorithm formulates the assembly problem as finding non-overlapping sets in an interval graph spanned over the input fragments. The algorithm exhibits lower computational complexity compared with a popular puzzle-solving approach based on minimal spanning trees.

Digital audio recordings are one of the key types of evidence used in law enforcement proceedings. As a result, the development of reliable techniques for forensic analysis of such recordings is of principal importance. One of the main problems in forensic analysis is source attribution, i.e., verifying whether a certain recording was acquired with a given device. While this problem has been widely studied for other types of multimedia signals, there are a very few techniques for audio recordings. Moreover, reported evaluation results were obtained from extremely small data sets on the order of a dozen devices. The goal of this paper is to propose a new feature set, the band energy difference (BED) descriptor, for source attribution of digital speech recordings. We demonstrate that a frequency response curve extracted from sample recordings can serve as a robust fingerprint that carries significant discriminative power and can characterize the recording device. We study two sub-problems of source attribution: 1) identification of a recording device among a list of possible candidates (device identification) and 2) confirming that a suspected device has indeed been used to acquire the recording in question (device verification). For our evaluation, we prepared two novel data sets: a controlled-conditions data set with 31 devices and an uncontrolled-conditions data set with 141 devices. Our experimental evaluation demonstrates that the proposed BED descriptor is effective for both device identification and verification. In the former task, we reached an accuracy of over 96%. In the latter, we obtained a high true positive rate of 89% while maintaining a fixed low false positive rate of 1%.

Accurate unsupervised tampering localization is one of the most challenging problems in digital image forensics. In this paper, we consider a photo response non-uniformity analysis and focus on the detection of small forgeries. For this purpose, we adopt a recently proposed paradigm of multi-scale analysis and discuss various strategies for its implementation. First, we consider a multi-scale fusion approach, which involves combination of multiple candidate tampering probability maps into a single, more reliable decision map. The candidate maps are obtained with sliding windows of various sizes and thus allow to exploit the benefits of both the small- and large-scale analyses. We extend this approach by introducing modulated threshold drift and content-dependent neighborhood interactions, leading to improved localization performance with superior shape representation and easier detection of small forgeries. We also discuss two novel alternative strategies: a segmentation-guided approach, which contracts the decision statistic to a central segment within each analysis window and an adaptive-window approach, which dynamically chooses analysis window size for each location in the image. We perform extensive experimental evaluation on both synthetic and realistic forgeries and discuss in detail practical aspects of parameter selection. Our evaluation shows that the multi-scale analysis leads to significant performance improvement compared with the commonly used single-scale approach. The proposed multi-scale fusion strategy delivers stable results with consistent improvement in various test scenarios.

In this paper we propose to use maximal entropy random walk on a graph for tampering localization in digital image forensics. Our approach serves as an additional post-processing step after conventional sliding-window analysis with a forensic detector. Strong localization property of this random walk will highlight important regions and attenuate the background - even for noisy response maps. Our evaluation shows that the proposed method can significantly outperform both the commonly used threshold-based decision, and the recently proposed optimization-based approach with a Markovian prior.

In this paper, we propose a novel method for generating visually appealing two-dimensional (2D) barcodes that resemble meaningful images to human observers. The technology of 2D barcodes, currently dominated by quick response codes, is widely adopted in many applications, including product tracking, document management, and general marketing. Such barcodes typically lack user friendly appearance and do not convey any visual significance to human observers. The proposed method addresses this problem by allowing 2D barcodes to resemble an arbitrary image or a logo. Our method is based on a generalization of the Kuznetsov-Tsybakov problem that served as a foundation for wet paper codes, commonly adopted in digital steganography. We introduce weaker statistical constraints to obtain additional flexibility allowing the barcode to assume the appearance of an arbitrary pattern. This paper provides the theoretical analysis of the proposed coding framework and a practical algorithm for rapid approximation of the optimal code. We also discuss the introduction of error correction capabilities, and experimentally evaluate a prototype implementation in a smartphone-based acquisition scenario.

A sliding window-based analysis is a prevailing mechanism for tampering localization in passive image authentication. It uses existing forensic detectors, originally designed for a full-frame analysis, to obtain the detection scores for individual image regions. One of the main problems with a window-based analysis is its impractically low localization resolution stemming from the need to use relatively large analysis windows. While decreasing the window size can improve the localization resolution, the classification results tend to become unreliable due to insufficient statistics about the relevant forensic features. In this paper, we investigate a multi-scale analysis approach that fuses multiple candidate tampering maps, resulting from the analysis with different windows, to obtain a single, more reliable tampering map with better localization resolution. We propose three different techniques for multi-scale fusion, and verify their feasibility against various reference strategies. We consider a popular tampering scenario with mode-based first digit features to distinguish between singly and doubly compressed regions. Our results clearly indicate that the proposed fusion strategies can successfully combine the benefits of small-scale and large-scale analyses and improve the tampering localization performance.

This paper deals with the design of a practical self-recovery mechanism for lossy compressed JPEG images. We extend a recently proposed model of the content reconstruction problem based on digital fountain codes to take into account the impact of emerging watermark extraction and block classification errors. In contrast to existing methods, our scheme guarantees a high and stable level of reconstruction quality. Instead of introducing reconstruction artifacts, emerging watermark extraction errors penalize the achievable tampering rates. We introduce new mechanisms that allow for handling high-resolution and color images efficiently. In order to analyze the behavior of our scheme, we derive an improved model to calculate the reconstruction success probability. We introduce a new hybrid mechanism for spreading the reference information over the entire image, which allows to find a good balance between the achievable tampering rates and the computational complexity. Such an approach reduced the watermark embedding time from the order of several minutes to the order of single seconds, even on mobile devices.

Annotation watermarking is a technique that allows to associate content descriptions with digital images in a persistent and format independent manner. It is commonly used in medical applications and, hence, existing schemes have been designed to meet rigorous watermark transparency requirements. As a result, the effective capacity of such schemes is severely limited. In this paper, we present a new approach to annotation watermarking. We adopt the fountain coding paradigm and design a convenient watermark communication architecture which resembles a traditional packet network. Our approach allows for straightforward incorporation of content adaptivity, robustness against cropping and support for multiple data streams. In our study, we focus on high-capacity annotations and we assume different requirements with respect to the fidelity of the watermarked images. Our scheme is robust against lossy JPEG compression and cropping. This paper describes the principles of the proposed approach and presents the results of it’s experimental evaluation.

Intelligent monitoring is currently one of the most prominent research areas. Numerous aspects of such schemes need to be addressed by implementation of various modules covering a wide range of algorithms, beginning from video analytic modules, through quality assessment, up to integrity verification. The goal of this paper is to provide a brief overview of the most recent research results regarding various aspects of the video surveillance processing chain. Specifically, the paper describes a scheme for automatic recognition of the make and model of passing vehicles, the state-of-the-art in quality assessment for recognition tasks, and a system for verification of digital evidence integrity. Concluding remarks highlight the perspectives for further development of the described techniques, and the related research directions.

This paper presents a new model of the content reconstruction problem in self-embedding systems, based on an erasure communication channel. We explain why such a model is a good fit for this problem, and how it can be practically implemented with the use of digital fountain codes. The proposed method is based on an alternative approach to spreading the reference information over the whole image, which has recently been shown to be of critical importance in the application at hand. Our paper presents a theoretical analysis of the inherent restoration trade-offs. We analytically derive formulas for the reconstruction success bounds, and validate them experimentally with Monte Carlo simulations and a reference image authentication system. We perform an exhaustive reconstruction quality assessment, where the presented reference scheme is compared to five state-of-the-art alternatives in a common evaluation scenario. Our paper leads to important insights on how self-embedding schemes should be constructed to achieve optimal performance. The reference authentication system designed according to the presented principles allows for high-quality reconstruction, regardless of the amount of the tampered content. The average reconstruction quality, measured on 10000 natural images is 37 dB, and is achievable even when 50% of the image area becomes tampered.

In this paper we analyze the content reconstruction problem with the use of a revised erasure communication channel. Based on this approach, we propose a reconfigurable self-embedding system which can be adapted to different requirements. Our approach eliminates two major problems with the design of efficient content reconstruction algorithms and allows for theoretical analysis of the reconstruction performance. The presented theoretical results are verified using Monte Carlo simulations. The proposed scheme is experimentally evaluated in a number of possible configurations, and allows to achieve high reconstruction quality even with high tampering rates.

In this paper we address the issue of the trade-off between the tampering rate and the reconstruction quality of image authentication systems. We adopt the fountain coding paradigm and design an adaptive content reconstruction scheme. The scheme conforms the reconstruction quality of individual image fragments both to the local texture properties and to the specified requirements. Experimental evaluation confirms that a framework based on this approach is a valid and convenient model of the performance of the considered reconstruction problem.

In this study we address the problem of an image quality trade-off that can be observed when dealing with content reconstruction schemes based on self-embedding. We derive two models for the estimation of optimal system parameters and the optimization of the overall image quality. This goal is achieved by balancing the distortions of a different nature that affect the resulting images. The performance of the derived models is verified with an accurate reference model and compared to traditional parameter selection strategies. The models are based on basic image features only and allow for rapid prediction of the best values for system parameters in a fully automatic manner.

Multimedia files often contain fragments with sensitive content that should not be visible to everyone. Such content is usually censored prior to distribution. We propose a technique that allows to blur selected fragments of the image while retaining details necessary for original appearance reconstruction in the image itself. The information is retained by means of a digital watermark. The protected image can be viewed in ordinary viewers and a dedicated decoder is required for content restoration. The decoder ensures that only authorized recipients are able to see the protected fragments. We evaluate the proposed approach in terms of protection capacity and reconstruction quality also under lossy compression. The results show that high image quality is achievable and the proposed scheme might be beneficial for certain applications. Moderate complexity of the decoding process makes it feasible for real-time usage scenarios.

If dynamic bandwidth-guaranteed connections between distributed services (e.g., grid services) are provisioned through a centralized system, the policy to serve connection requests might heavily impact both the success in and the time required for setting up user services (e.g., grid-enabled applications). In this paper, the implementation of a batch queue in the centralized system is proposed. By implementing different service policies for the queued requests, connections and, in consequence, user services can be set up with different guarantees. In this study, a bulk-service policy is proposed and implemented to maximize connection set up success. The experimental evaluation results show that the utilization of the proposed policy brings advantages in terms of percentage of accepted connection requests as the number of requests served in one batch increases. Moreover, the achieved improvement does not impact the time required to set up the connections because of the specific LSP set up procedures implemented in the utilized commercial routers.