Source Camera Identification Research Articles

Source camera identification for re-compressed images: A model perspective based on tri-transfer learning

Source Camera Identification (SCI) achieves high accuracy on matching identification, in which the training and testing sample sets are derived from the same statistical distribution. However, in practice the training and testing sets, namely, the source and test domains, may consist of digital images that are double compressed by various software and applications with different quantization tables. Unfortunately, existing methods are inadequate in performance under such circumstances, such that we aim to find an algorithm that can fill the gap between the training and testing sets. In this work, we propose an algorithm, tri-transfer Learning (TTL), which is a cross-pollination of transfer learning and tri-training. For TTL, the transfer learning module transfers the knowledge learned from the training sets to improve the identification performance on testing. Compared with other methods, TTL uses a semi-supervised approach requiring only a small number of training samples and has better performance than other methods. The tri-training module, which is a variation of the co-training, facilitates knowledge transferring by assigning pseudo-labels to unlabelled instances and adds target instances with labels to the training set in batches. Combining the two modules, our framework can gain higher efficiency and performance than other state-of-art methods on mismatched camera model identification which is supported by experiments based on the open-source Dresden Image Database.

Blind Image Clustering for Camera Source Identification via Row-Sparsity Optimization

Given a set of images with the number of cameras providing those images unknown, how to blindly identify the sources of the images has been a critical problem in digital forensics. Although state-of-the-art methods have achieved impressive results, they have failed at suppressing outliers. When they deal with a noisy dataset, the performance is significantly degraded. To address this issue, we propose an optimization approach with sparsity constraints to simultaneously handle the how-many subproblem ( i.e. , the number of cameras) and the which-from-which subproblem ( i.e. , the image–camera relationship). In our approach, we first formulate the blind camera source clustering as a row-sparsity optimization problem, in which the representation errors are minimized and the outliers caused by noisy features are suppressed. Then, a new two-stage refinement method based on inter- and the intra-class differences is proposed to achieve a more accurate estimation of the number of cameras. Because strong sparsity constraints have been adopted and the interactive relationship among data points can be fully explored to distinguish the images originated from different cameras, the proposed method can effectively handle outliers. Extensive experiments on the popular Dresden dataset show that the proposed method outperforms existing methods in both identification accuracy and efficiency.

Deep siamese network for limited labels classification in source camera identification

Source Camera Identification is a well-known digital forensic challenge of mapping an image to its authentic source. The current state-of-the-art provides a number of successful and efficient solutions to this problem. However, in almost all such existing techniques, a sufficiently large number of image samples is required for pre-processing, before source identification. Limited labels classification is a realistic scenario for a forensic analyst where s/he has access only to a few labelled training samples, available for source camera identification. In such contexts, where obtaining a vast number of image samples (per camera) is infeasible, correctness of existing source identification schemes, is threatened. In this paper, we address the problem of performing accurate source camera identification, with a limited set of labelled training samples, per camera model. We use a few shot learning technique known as deep siamese network here, and achieve significantly improved classification accuracy than the state–of–the–art. Here, the main principle of operation is to form pairs of samples from the same camera models, as well as from different camera models, to enhance the training space. Subsequently, a deep neural network is used to perform source classification. We perform experiments on traditional camera model identification, as well as intra–make and intra–device source identification. We also show that our proposed methodology under limited labels scenario, is robust to image transformations such as rotation, scaling, compression, and additive noise.

A Survey of Deep Learning-Based Source Image Forensics.

Image source forensics is widely considered as one of the most effective ways to verify in a blind way digital image authenticity and integrity. In the last few years, many researchers have applied data-driven approaches to this task, inspired by the excellent performance obtained by those techniques on computer vision problems. In this survey, we present the most important data-driven algorithms that deal with the problem of image source forensics. To make order in this vast field, we have divided the area in five sub-topics: source camera identification, recaptured image forensic, computer graphics (CG) image forensic, GAN-generated image detection, and source social network identification. Moreover, we have included the works on anti-forensics and counter anti-forensics. For each of these tasks, we have highlighted advantages and limitations of the methods currently proposed in this promising and rich research field.

An experimental estimate of the impact produced on PNU by new generation video codecs

The resolution of video cameras has increased considerably in recent years leading to a new generation of codecs with higher compression rates, such as H.265, VP9 and AV1. However, compression can heavily affect the noise present in each frame. Other encoder specific features, like the intra-frame prediction for H.265 and AV1, can flatten the pixel non-uniformity (PNU) noise. We implemented a test-bed to establish whether source camera identification can still be achieved using the PNU when videos are encoded in those formats. The experiments proved that the results are less accurate than those obtained from videos encoded with traditional codecs. Nevertheless, the most original result achieved was the methodology we designed. In order to avoid hidden artifacts the source videos were initially acquired as a raw data stream from a dedicated embedded system and then recorded with an open-source YUV encoder.

Source Camera Identification With Dual-Tree Complex Wavelet Transform

Sensor pattern noise (SPN) extraction is a critical stage of the sensor based source camera identification (SCI). However, the quality of the extracted SPN with the traditional discrete wavelet transform (DWT) based method is poor around strong edges and along with the image border. To fill this gap, we propose a dual tree complex wavelet transform (DTCWT) based method to extract the SPN from a given image, which achieves better performance in the area around strong edges. Furthermore, symmetric boundary extension instead of the periodized boundary extension is used for enhancing the quality of SPN along with the image border. Extensive experimental results on both synthetic noisy images and real-world photographs clearly demonstrate the superior SCI performance of the proposed method over state-of-the-arts. Moreover, the proposed method also shows potential in the application of image tampering localization.

Source Camera Identification Based on Coupling Coding and Adaptive Filter

Source Camera Identification (SCI) has been playing an important role in the security field for decades. With the development of Deep Learning, the performance of SCI has been noteworthily improved. However, most of the proposed methods are forensic only for a single camera identification category, e.g., the camera model identification. For exploiting the coupling between different camera categories, we present a new coding method. That is, we apply the multi-task training method to regress the categories, namely, to classify brands, models and devices synchronously in a single network. Different from the common multi-task method, we obtain the multi-class classification result by just one single label classification. To be specific, we classify the categories in a progressive way that the parent category classification result will be used in the child category classification (a detailed explanation will be given later in the main context). Also, by appropriately increasing the redundancy of the coding method for classifying new camera categories, the training time can be greatly reduced. To better extract camera attributes, we propose an adaptive filter. Additionally, we propose an auxiliary classifier that only focuses on the camera model re-classification, due to the low performance of the main classifier on certain models. Lastly, the extensive experiments show that our methods have a better performance than other existing methods.

A unified framework of source camera identification based on features

Source camera identification, which aims at identifying the source camera of an image, has attracted a wide range of attention in the field of digital image forensics recently. Many approaches to source camera identification have been proposed by extracting some image features. However, most of these methods only focused on extracting features from the single artifact of the camera left on the captured images and ignored other artifacts that may help improve final accuracy. Therefore, in this paper, we propose a feature-based framework for source camera identification, which first captures various pure camera-specific artifacts through preprocessing and residual calculation, then extracts discriminative features through image transform, and finally reduces the algorithm complexity through feature reduction. Based on the framework, a novel source camera identification method is proposed, which can identify different camera brands, models and individuals with high accuracy. A large number of comparative experiments show that the proposed method outperforms the state-of-the-art methods.

Camera model identification using a deep network and a reduced edge dataset

Today, the importance of digital images as a medium for social communication is growing rapidly. Sometimes, an image needs to be authenticated by verifying its source camera model or device. Recently, deep networks have become very successful at visual pattern recognition. With this motivation, several investigators have explored the possibility of using convolutional neural networks (CNNs) for camera source identification. In this paper, we use selective preprocessing, instead of a indiscriminate one, in order not to hinder the CNN’s strong ability to learn useful features for this kind of forensic task. To generate a consistent and balanced dataset, we limit the maximum number of original images to 200 per camera model, and we discard vertically taken images. Using a relatively simple deep network structure, the proposed method achieved a better prediction accuracy—95.0%—than GoogleNet and other existing methods. Also, challenging camera models such as the Sony DSC H50 and W170 can be classified with the quite high prediction accuracies of 87.9% and 83.1%, respectively.

A PUF-Based Data-Device Hash for Tampered Image Detection and Source Camera Identification

With the increasing prevalent of digital devices and their abuse for digital content creation, forgeries of digital images and video footage are more rampant than ever. Digital forensics is challenged into seeking advanced technologies for forgery content detection and acquisition device identification. Unfortunately, existing solutions that address image tampering problems fail to identify the device that produces the images or footage while techniques that can identify the camera is incapable of locating the tampered content of its captured images. In this paper, a new perceptual data-device hash is proposed to locate maliciously tampered image regions and identify the source camera of the received image data as a non-repudiable attestation in digital forensics. The presented image may have been either tampered or gone through benign content preserving geometric transforms or image processing operations. The proposed image hash is generated by projecting the invariant image features into a physical unclonable function (PUF)-defined Bernoulli random space. The tamper-resistant random PUF response is unique for each camera and can only be generated upon triggered by a challenge, which is provided by the image acquisition timestamp. The proposed hash is evaluated on the modified CASIA database and CMOS image sensor-based PUF simulated using 180 nm TSMC technology. It achieves a high tamper detection rate of 95.42% with the regions of tampered content successfully located, a good authentication performance of above 98.5% against standard content-preserving manipulations, and 96.25% and 90.42%, respectively, for the more challenging geometric transformations of rotation (0 ~ 360°) and scaling (scale factor in each dimension: 0.5). It is demonstrated to be able to identify the source camera with 100% accuracy and is secure against attacks on PUF.

PRNU based source camera attribution for image sets anonymized with patch-match algorithm

Patch-Match is an efficient algorithm used for structural image editing and available as a tool on popular commercial photo-editing software. The tool allows users to insert or remove objects from photos using information from similar scene content. Recently, a modified version of this algorithm was proposed as a counter-measure against Photo-Response Non-Uniformity (PRNU) based Source Camera Identification (SCI). The algorithm can provide anonymity at a great rate (97%) and impede PRNU based SCI without the need of any other information, hence leaving no-known recourse for the PRNU-based SCI. In this paper, we propose a method to identify sources of the Patch-Match-applied images by using randomized subsets of images and the traditional PRNU based SCI methods. We evaluate the proposed method on two forensics scenarios in which an adversary makes use of the Patch-Match algorithm and distorts the PRNU noise pattern in the incriminating images she took with his camera. Our results show that it is possible to link sets of Patch-Match-applied images back to their source camera even in the presence of images that come from unknown cameras. To our best knowledge, the proposed method represents the very first counter-measure against the usage of Patch-Match in the digital forensics literature.

Semantic ear feature reduction for source camera identification

Semantic ear feature reduction for source camera identification

Achieving efficient source camera identification on Hadoop

Hadoop is a software framework allowing for the possibility of coding distributed applications starting from a MapReduce algorithm with very low programming efforts. However, the performance of the implementations resulting from such a straightforward approach are often disappointing. This may happen because a vanilla implementation of a MapReduce distributed algorithm often suffers of some performance bottlenecks that may compromise the potential of a distributed system. As a consequence of this, the execution times of the considered algorithm are not up to the expectations. In this paper, we present the work we have done for efficiently engineering, on Apache Hadoop, a reference algorithm for the Source Camera Identification problem (i.e., determining the particular digital camera used for taking a given image). The algorithm we have chosen is the algorithm by Lukas et al.. A first implementation has been obtained in a small amount of time using the default facilities available with Hadoop. However, its performance, analyzed using a cluster of 33 PCs, was very unsatisfactory. A careful profiling of this code revealed some serious performance issues targeting the initial steps of the algorithm and resulting in a bad usage of the cluster resources. Several theoretical and practical optimizations were then tried, and their effects were measured by accurate experimentations. This allowed for the development of alternative implementations that, while leaving unaltered the original algorithm, were able to better use the underlying cluster resources as well as of the Hadoop framework, thus allowing for much better performance and reduced energy requirements than the original vanilla implementation.

Isolating lens effects from source camera identification using sensor pattern noise

ABSTRACTLens aberrations can be used to determine the provenance of camera optics in an image. However, this method does not necessarily identify the camera itself, since lens systems are often interchanged. A method, first published by Lukàš, Fridrich and Goljan, proposed using Fixed Pattern Noise to link a photo to an image sensor using sensor pattern noise. However, their model did not fully account for lens artefacts or temperature. In our work, we have applied standard image processing theory and an understanding of the geometric properties of light and sensor dark current to continue the isolation of artefacts within the sensor pattern noise model. We use three image sensors to take images using six integrated lenses and prepare reference patterns based on the lens for each camera. We then repeat our experiments with a pinhole lens of our design before comparing the correlation energy contained within each. We show that majority of the unique signal power used to correlate an image to a specific camera is related to the photo response non-uniformity noise; however, additional signal power can also be attributed to the optical system of the camera as well as the sensor’s dark current.

Joint statistics matching for camera model identification of recompressed images.

Source camera identification has been well studied in laboratory environment where the training and test samples are all original images without recompression. However, image compression is quite common in the real world, when the training and test images are double JPEG compressed with different quantization tables, the identification accuracy of existing methods decreases dramati- cally. To address this challenge, we propose a novel iterative algorithm namely joint first and second order statistics matching (JSM) to learn a feature projection that projects the training and test fea- tures into a low dimensional subspace to reduce the shift caused by image recompression. Inspired by transfer learning, JSM aims to learn a new feature representation from original feature space by simultaneously matching the first and second order statistics between training and test features in a principled dimensionality reduction procedure. After the feature projection, the divergence between training and test features caused by recompression is reduced while the discriminative properties are preserved. Extensive experiments on public Dresden Image Database verify that JSM significantly outperforms several state-of-the-art methods on camera model identification of recompressed images.

Ensemble classifier based source camera identification using fusion features

Source camera identification, which means identifying the camera source of a given image, has become one of the most important branches of digital image forensics. In order to improve the detection accuracy, the feature dimensions used in existing methods are increasing, and consequently Support Vector Machine (SVM) seems no longer applicable. In this paper, an ensemble classifier is introduced into to source camera identification, which uses the fusion features to capture software-related, hardware-related, and statistical characteristics left on the images by digital camera. Experimental results indicate that the proposed method can achieve near 100% accuracy for camera brand and model identification, and also outperforms the baseline methods in identifying different camera individuals.

Source camera identification via low dimensional PRNU features

Identifying the source of digital images is the key task in the community of image forensics. Sensor pattern noise dominantly serves as an intrinsic fingerprint or feature for dealing with the problem of source camera identification. However, how to decrease the dimensionality of the pattern noise while guaranteeing the detection power remains a hot topic. The goal of this paper is to investigate the problem of source camera identification for natural images in JPEG format. By considering the image texture, we propose to design a new classifier with adopting a weight function, leading to the remarkable reduction of the feature dimensionality. In the extensive experiments, it is verified that our proposed algorithm performs comparably with the prior art. Besides, the robustness of the proposed classifier is also evaluated when the query images are attacked by post-processing techniques such as JPEG compression, noise adding, noise removing and image cropping.

Wavelet energy feature based source camera identification for ear biometric images

In this paper a source camera identification algorithm for ear biometric images has been proposed based on tunable filter bank as a feature extractor. Maintaining the frequency selectivity property, distinct features are extracted by this filter bank, based on a half-band polynomial of 14th order. With the help of four ear databases, it is demonstrated that tunable filter bank based features correctly identify the sources of ear images with an average accuracy of 99.25% when there are limited number of camera sources available. It is also shown that accuracy would fall when significantly large number of cameras are introduced to acquire ear images. Depending on the experimental results, it can be well concluded that tunable filter bank based feature, apart from its recognition performance, is also a promising candidate to support forensic validation of camera source.

An empirical cross-validation of denoising filters for PRNU extraction

The present work is an empirical study to check the importance of widely used PRNU extraction de-noising filters at different stages of source camera identification procedure. The proposed work is unique in the sense that it gives an idea about the choice of the appropriate de-noising filters at the time of PRNU extraction for formation of unique identification pattern of digital camera and noise residue extraction of query image. Also in this work, we have suggested the best values of σ (noise variance) for formation of unique identification pattern of digital camera and noise residue extraction of query image (based on empirical observations). This study has been performed to check that which part (camera unique identification pattern, noise residue, enhancement methods, and value of σ) mostly dominates the performance of source camera identification.

Camera Source Identification of Digital Images Based on Sample Selection

With the advent of the Information Age, the source identification of digital images, as a part of digital image forensics, has attracted increasing attention. Therefore, an effective technique to identify the source of digital images is urgently needed at this stage. In this paper, first, we study and implement some previous work on image source identification based on sensor pattern noise, such as the Lukas method, principal component analysis method and the random subspace method. Second, to extract a purer sensor pattern noise, we propose a sample selection method to improve the random subspace method. By analyzing the image texture feature, we select a patch with less complexity to extract more reliable sensor pattern noise, which improves the accuracy of identification. Finally, experiment results reveal that the proposed sample selection method can extract a purer sensor pattern noise, which further improves the accuracy of image source identification. At the same time, this approach is less complicated than the deep learning models and is close to the most advanced performance.