Forensic Analysis for Source Camera Identification from EXIF Metadata.

Photo Response Non–Uniformity (PRNU) is one of the most effective fingerprints used to detect the source camera of an image. Image Anonymization on the other hand, is a task of fooling the source camera identification, in order to protect the user’s anonymity in sensitive situations involving whistleblowers, social activists etc. To protect the privacy of users especially over the web, image anonymization is of huge importance. Counter–Forensic attacks on source camera identification try to make an image anonymous by nullifying the detection techniques. For almost every counter–forensic source camera identification attack, anti–counter attacks are being designed and hence there is a need to either strengthen the previous counter–forensic attacks or design a new attack altogether. In this work, we propose a new counter–forensic attack to source camera identification, using the Universal Wavelet Relative Distortion function designed for steganography. The main principle behind Universal Wavelet Relative Distortion is to embed changes in an image in regions such as textures or noisy parts which are crucial to source camera identification. We show through our experiments, when a random bit–string is inserted recursively in an image, the correlation strength of the noise residual based source camera identification gets significantly weak and such methods fail to map the source camera of the image under question. In the proposed method, the visual quality of the modified image is not changed, which makes our method a strong solution to image anonymization.

Recent research found that attacks based on morphed face images, i.e., morphing attacks, pose a severe security risk to face recognition systems. A reliable morphing attack detection from a single face image remains a research challenge since cameras and morphing techniques used by an attacker are unknown at the time of classification. These issues are commonly overseen while many researchers report encouraging detection performance for training and testing morphing attack detection schemes on images obtained from a single face database employing a single morphing algorithm. In this work, a morphing attack detection system based on the analysis of Photo Response Non-Uniformity (PRNU) is presented. More specifically, spatial and spectral features extracted from PRNU patterns across image cells are analyzed. Differences of these features for bona fide and morphed images are estimated during a threshold-selection stage using the Dresden image database which is specifically built for PRNU analysis in digital image forensics. Cross-database evaluations are then conducted employing an ICAO compliant subset of the FRGCv2 database and a Print-Scan database which is a printed and scanned version of said FRGCv2 subset. Bona fide and morphed face images are automatically generated employing four different morphing algorithms. The proposed PRNU-based morphing attack detector is shown to robustly distinguish bona fide and morphed face images achieving an average D-EER of 11.2% in the best configuration. In scenarios where image sources and morphing techniques are unknown, it is shown to significantly outperform other previously established morphing attack detectors. Finally, the limitations and potential of the approach are demonstrated on a dataset of printed and scanned bona fide and morphed face images.

Sensor noise caused by photo response non-uniformity (PRNU) has been widely accepted as a reliable fingerprint for source camera identification (SCI). In the sensor based camera identification methods, the extracted noise from a test image needs to be absolutely synchronized with the camera fingerprint. However, for the images captured under digital zoom, this requirement is not satisfied, which is very common for smartphone shooting environment due to no optical zoom available. Here, through carefully studying about the effect of the digital zoom on the sensor-based SCI, we found several characteristics in this case, and proposed a novel metric to enhance the SCI performance for images taken with digital zoom. Experimental results show that the proposed metric outperforms the benchmark metric for source smartphone identification on digital zoomed images by evaluating the ROC performance.

Photo-response non-uniformity (PRNU) noise has been well established as a reliable fingerprint of imaging sensors for source camera identification and other forensic applications. In this paper, we introduce a novel denoising method for PRNU noise extraction that considerably outperforms the existing techniques. The rationale is to use as little as one adjacent pixel in spatial domain filtering to suppress the pixel-to-pixel correlation in the estimation of the (supposedly white) PRNU noise. Experimental results are presented based on an image database of ten cameras of different models and makes. These results are embodied in the receiver operating characteristic (ROC) of source camera identification.

Source identification is one of the most critical problems in the field of multimedia forensics. In the last decade, researchers have been studying and improving in this field. Photo Response Non-Uniformity is one of the unique noise patterns that is being used to match a media to its originating device. Utilizing the noise patterns with machine learning algorithms has been the focus of research in recent years. Therefore, a systematic review is needed to present the latest contributions in this field. This systematic review focuses on the published work from 2015 to 2021 in source identification using noise patterns in machine learning-based systems. The results of the review indicate that a benchmark should be proposed and used to fairly compare past and future methods. Moreover, a minimum number of devices used for evaluating a model should be set by the research community in order to accurately evaluate the accuracy of the model in real-life situations.

Source camera identification can verify whether two videos were shot by the same device, which is of great significance in multimedia forensics. Most existing identification methods use convolutional neural networks to learn sensor noise patterns to identify the source camera in closed forensic scenarios. While these methodologies have achieved remarkable results, they are nonetheless constrained by two primary challenges: (1) the interference of semantic information and (2) the incongruity in feature distributions across different datasets. The former will interfere with the extraction of effective features of the model. The latter will cause the model to fit the characteristic distribution of the training data and be sensitive to unseen data features. To address these challenges, we propose a novel source camera identification framework that determines whether a video was shot by the same device by obtaining similarities between source camera features. Firstly, we extract video key frames and use the integral image to optimize the smoothing blocks selection algorithm of inter-pixel variance to remove the interference of video semantic information. Secondly, we design a residual neural network fused with a constraint layer to adaptively learn video source features. Thirdly, we introduce a triplet loss metric learning strategy to optimize the network model to improve the discriminability of the model. Finally, we design a multi-dimensional feature vector similarity fusion strategy to achieve highly generalized source camera recognition. Extensive experiments show that our method achieved an AUC value of up to 0.9714 in closed-set forensic scenarios and an AUC value of 0.882 in open-set scenarios, representing an improvement of 5% compared to the best baseline method. Furthermore, our method demonstrates effectiveness in the task of deepfake detection.

In today’s digital landscape, YouTube’s comment sections play an essential role in shaping public opinion and discussions. However, coordinated groups often exploit these platforms to spread misinformation and distort conversations. This study introduces a method to detect and quantify such anomalous behavior on YouTube channels by analyzing two key components: commenter behaviors and engagement patterns. Focusing on channels concerning news outlets, geopolitics, and the military, we draw from a dataset with 71 channels, 642,952 videos, 12,425,587 commenters, 123,882,200 comments, 139,985,870 subscribers, and 83,396,188,807 views. Our approach leverages key engagement indicators, such as comments, views, and subscriber counts, alongside an analysis of frequent commenter interactions to identify anomalous patterns. We apply unsupervised techniques, including kernel density estimation (KDE) and Gaussian mixture model (GMM), to assign a score reflecting anomalous commenter behavior. To provide a comprehensive measure of this behavior, we combine commenter and engagement scores both at the feature level and at the output level. At the feature level, we employ cosine similarity and principal component analysis (PCA). At the output level, we propose three scoring methods: harmonic mean (HM), weighted average with interaction term (WAIT), and agreement-weighted maximum (AWM). The resulting score, normalized between 0 and 1, indicates the level of anomalous activity on each channel. We validate our results by comparing detected anomalous channels with actual suspended channels from YouTube, confirming the model’s real-world applicability. By offering a quantifiable measure of this activity, our method helps preserve the integrity of YouTube discussions.

• Constructed the largest multi-category dataset for tunnel lining defect detection following rigorous data collection, analysis, and processing. • Proposed an improved KNet-based segmentation model that is validated to outperform multiple composite algorithms under a consistent transfer learning strategy. • Analyzed the model's decision-making process and performance factors in depth via progressive visual interpretability and multi-index joint analysis. • Developed a desktop application for the Central Yunnan Water Diversion Project to facilitate the efficient detection and identification of lining defects in experimental sections. During tunnel construction and operation, the timely detection and treatment of lining defects are critical to ensuring the safety and stability of tunnel engineering. To achieve fast and accurate detection of multi-category tunnel lining defects from image data, this paper proposes an improved model based on the KNet segmentation algorithm, integrated with the Swin Transformer and UPerNet models. First, a large-scale multi-category image dataset of tunnel lining defects is constructed by collecting and organizing data from relevant literature and public datasets. Second, the pre-trained KNet-enhanced model is trained on this dataset using transfer learning to improve detection accuracy and generalization performance. The proposed KNet-UPerNet-SwinL model achieves an average improvement of approximately 16.94% in mean Intersection over Union (mIoU) compared to other classical semantic segmentation methods, demonstrating superior segmentation accuracy and stability. In addition, an in-depth discussion of the model’s decision-making process and the factors influencing its performance is provided through visually interpretable progressive feature analysis and multi-index joint analysis at the application level. Finally, to facilitate practical application, the model is converted into the Open Neural Network Exchange (ONNX) universal format, and a desktop application for tunnel lining defect detection is developed. The detection and recognition of lining defects are successfully implemented in the experimental sections of the Central Yunnan Water Diversion Project, achieving remarkable segmentation results. In summary, this study provides a comprehensive solution that offers practical technical and tool support for multi-category defect detection in tunnel linings and holds significant importance for the safety and stability of tunnel engineering.

Sensor pattern noise (SPN) extracted from digital images has been proved to be a unique fingerprint of digital camera. However, sensor pattern noise can be contaminated largely in frequency domain by image detail from scene according to Li's work and non-unique artifacts of on-sensor signal transfer, sensor design, color interpolation according to Chen et al's work, the source camera identification performance based on SPN needs to be improved especially for small image block. Motivated by their works, in order to lessen the effect of these contaminations, the unique SPN fingerprint for identifying one specific camera is assumed to be a white noise which has a flat frequency spectrum, so the SPN extracted from an image is whitened first to have a flat frequency spectrum, then inputted to the mixed correlation detector. Source camera identification is the detection of the existence of the camera reference SPN in the SPN extracted from a single image. Compared with the correlation detection approach and Li's model based approaches on 7 cameras, 1400 photos totally, each camera is responsible for 200, the experimental results show that the proposed mixed correlation detection enhances the receiver operating characteristic (ROC) performance of source camera identification, especially greatly raises the detection rate (true positive rate) in the case of trustworthy identification which is with a low false positive rate. For example, the proposed mixed correlation detection raises the true positive rate from 78% to 93% at zero false positive rate on image blocks of 256x256 pixels cropped from the center of the 1400 photos. The proposed mixed correlation detection also has large advantage to resist JPEG compression with low quality factor. Fridrich's group has proposed two reference SPN extraction methods which are the noise residues averaging and the maximum likelihood estimation method. They are compared from the aspect of ROC performance associated with the correlation detection and mixed correlation detection respectively. It is observed that the combination of mixed correlation detection and reference SPN extraction method of noise residues averaging achieves the best performance. We also demonstrate an image management application of the proposed SPN detection method for the news agent. It shows that the detection method discriminates the positive samples from a large number of negative samples very well on image bock size of 512×512 pixels.

We endeavor to develop a novel deep learning architecture tailored specifically for the analysis and classification of tongue features, including color, shape, and coating. Unlike conventional methods based on architectures like VGG or ResNet, our proposed method aims to address the challenges arising from their extensive size, thereby mitigating the overfitting problem. Through this research, we aim to contribute to the advancement of techniques in tongue feature recognition, ultimately leading to more precise diagnoses and better patient rehabilitation in Traditional Chinese Medicine (TCM). In this study, we introduce TGANet (Tongue Feature Attention Network) to enhance model performance. TGANet utilizes the initial five convolutional blocks of pre-trained VGG16 as the backbone and integrates an attention mechanism into this backbone. The integration of the attention mechanism aims to mimic human cognitive attention, emphasizing model weights on pivotal regions of the image. During the learning process, the allocation of attention weights facilitates the interpretation of causal relationships in the model's decision-making. Experimental results demonstrate that TGANet outperforms baseline models, including VGG16, ResNet18, and TSC-WNet, in terms of accuracy, precision, F1 score, and AUC metrics. Additionally, TGANet provides a more intuitive and meaningful understanding of tongue feature classification models through the visualization of attention weights. In conclusion, TGANet presents an effective approach to tongue feature classification, addressing challenges associated with model size and overfitting. By leveraging the attention mechanism and pre-trained VGG16 backbone, TGANet achieves superior performance metrics and enhances the interpretability of the model's decision-making process. The visualization of attention weights contributes to a more intuitive understanding of the classification process, making TGANet a promising tool in tongue diagnosis and rehabilitation.

The global increase in endometrial cancer, including in Japan, and a shortage of pathologists and cytotechnologists have increased the diagnostic burden, emphasizing the need for an AI-based diagnostic support model that uses deep learning. We evaluated the clinical application of an improved AI-supported endometrial cytology model. Using YOLOv5x and YOLOv7 models evaluated by mean average precision (mAP), we compared two datasets-one annotated for both benign and malignant cell clusters, and one for malignant only. In addition, using the Two One-Sided Tests (TOST) procedure, we assessed the correlation between AI diagnostic accuracy and the level of difficulty perceived by human diagnosticians. Finally, we used Gradient-weighted Class Activation Mapping (Grad-CAM) to visualize and enhance the interpretability of the AI model's decision-making process. The YOLOv5x model with both benign and malignant annotations had the highest malignant mAP, 0.798, as compared with YOLOv7. The TOST analysis showed no significant difference in perceived diagnostic difficulty between cases that were correctly and incorrectly diagnosed by the AI model, indicating consistent AI accuracy regardless of case difficulty. Grad-CAM visualizations clarified the AI model's decision-making basis; in some cases, the model appeared to focus on regions that differed from those typically attended to by human diagnosticians. The AI support model showed high and consistent accuracy in endometrial cytological analysis, regardless of diagnostic difficulty as perceived by human diagnosticians. Grad-CAM visualizations revealed diagnostic patterns, and the AI occasionally focused on regions different from those emphasized by human diagnosticians. This study advanced a real-time microscope-integrated AI system toward clinical application.

The proliferation of AI-generated imagery poses escalating challenges for multimedia forensics, yet many existing detectors depend on assumptions about the internals of specific generative models, limiting their cross-model applicability. We introduce a self-supervised approach for detecting AI-generated images that leverages camera metadata-specifically exchangeable image file format (EXIF) tags-to learn features intrinsic to digital photography. Our pretext task trains a feature extractor solely on camera-captured photographs by classifying categorical EXIF tags (e.g., camera model and scene type) and pairwise-ranking ordinal and continuous EXIF tags (e.g., focal length and aperture value). Using these EXIF-induced features, we first perform one-class detection by modeling the distribution of photographic images with a Gaussian mixture model and flagging low-likelihood samples as AI-generated. We then extend to binary detection that treats the learned extractor as a strong regularizer for a classifier of the same architecture, operating on high-frequency residuals from spatially scrambled patches. Extensive experiments across various generative models demonstrate that our EXIF-induced detectors substantially advance the state of the art, delivering strong generalization to in-the-wild samples and robustness to common benign image perturbations.

Localizing image forgeries is one of the key topics in multimedia forensics. Among many image forgery localization techniques, the one based on the photo-response non-uniformity (PRNU) noise has attracted substantial attention because of its capability of localizing forgeries regardless of the type of forgery. However, despite the devoted efforts to improving the performance of PRNU-based forgery localization, there remain challenges to be overcome, especially for detecting subtle forgeries in PRNU-attenuated regions due to complex image content. In this work, we investigate the feasibility and effectiveness of convolutional neural networks (CNN) in predicting PRNU correlations under complex backgrounds for more accurate forgery localization. The experimental results on 20 cameras and 200 realistic forgery images show that significant improvement in correlation prediction and forgery localization can be achieved even with a light-weight CNN model. The robustness of different correlation predictors against JPEG compression is also evaluated.

The verification of Unmanned Aerial Systems (UAS) trajectory generation in uncertain environments is a critical challenge, particularly when using reinforcement learning (RL) methods. The unpredictability of RL models, together with the lack of formal verification standards for UAS and AI systems, complicates the safety and reliability of autonomous decision-making. This paper proposes a learning framework as a solution towards improving the verifiability of RL-based trajectory generation. We propose an incremental training process where the RL model is progressively trained in more complex and uncertain scenarios. By controlling the complexity of training environments, the framework can guide the system towards safer, more predictable behaviour, making it easier to verify its performance under varying conditions. This approach also helps reduce the risks associated with RL's non-deterministic nature by creating a better understanding of robustness and stability in the model's decision-making process. Additionally, we discuss the current means of certifying airborne software and the need for new standards and methodologies to ensure that learning-based systems can be rigorously evaluated for safety and operational readiness in dynamic, real-world environments.

Pneumonia is a serious respiratory infection that can lead to life-threatening complications, particularly in vulnerable populations such as infants and seniors. Early and accurate diagnosis is crucial for effective treatment and improved patient outcomes. This paper presents a novel deep learning framework for pneumonia detection using chest X-ray images, based on the Swin Transformer architecture. The model effectively captures both local and global contextual information from medical images, enabling robust feature representation. Trained and evaluated on the pediatric chest X-ray dataset curated by Kermany et al., the proposed approach incorporates comprehensive preprocessing and data augmentation techniques to enhance generalization capability. The system achieves a classification accuracy of 94%, outperforming conventional deep learning architectures including VGG16, ResNet50, DenseNet121, and EfficientNetB0. To address the critical need for transparency in medical AI, the framework integrates Explainable AI (XAI) methods, specifically Grad-CAM and EigenCAM, which provide visual explanations of the model's decision-making process by highlighting clinically relevant regions in lung images. Experimental results demonstrate that the proposed method not only delivers high diagnostic accuracy but also offers interpretable predictions, making it a promising decision-support tool for clinical practice and facilitating early pneumonia detection.