Image Forensics Dataset with Manipulation Masks
Year: 2026
Type: dataset
Abstract: 1. AbstractThe Image Forensics Dataset is designed to support research in digital image forensics, image forgery detection, and image manipulation localization. It consists of authentic images and synthetically manipulated images representing two common forgery techniques: copy-move forgery and image splicing.The dataset includes pixel-level manipulation masks for identifying forged regions and separate source-region masks for copy-move forgery. A metadata file provides a centralized record of dataset information, while the organized directory structure facilitates image retrieval, annotation access, and experimental use.The dataset can be used to develop and evaluate deep learning models for image authenticity classification, forgery detection, and pixel-level manipulation localization. It provides a resource for studying image manipulation detection and comparing classification and segmentation approaches in digital image forensics.2. Dataset ContentsThe dataset consists of three image categories and their associated annotations, where applicable.Authentic images: Original images without intentionally introduced manipulations. These images provide reference samples for image authenticity classification.Copy-move forgery: Images in which a region is copied and pasted into another location within the same image. Corresponding manipulation masks identify the pasted regions, while source masks identify the original regions used for copying.Image splicing: Images created by inserting content from one image into another. Corresponding masks identify the inserted regions in the resulting images.Metadata: The metadata.csv file stores dataset-related information in tabular form. Its specific fields depend on the metadata schema.Documentation: The README.txt file provides a location for dataset descriptions, usage instructions, and annotation conventions.3. Mask GenerationThe dataset provides segmentation masks to support supervised learning for image manipulation localization.Copy-move masks: Each mask identifies the region designated as manipulated after content has been copied and pasted within the same image. The corresponding source mask identifies the original region from which the content was copied. Maintaining this distinction allows researchers to study both the original and duplicated regions.Splicing masks: Each mask identifies the region designated as inserted into the target image from another image. These annotations enable models to learn spatial patterns associated with manipulated content.Authentic images: The authentic category contains images without intentionally introduced forgeries. No mask directory is included for this category in the current dataset structure.The exact mask encoding should be documented according to the actual files. For binary masks, a common convention is 0 for background and 255 for manipulated pixels, although other encodings may be used.4. Important Interpretation NoteThe dataset contains synthetically generated forgeries and automatically generated segmentation annotations. During annotation generation, YOLOv8 is used for object detection, while Segment Anything Model 2 (SAM 2) is used to produce object-level segmentation masks.Although these models facilitate object boundary identification, their outputs should not automatically be treated as manually verified forensic ground truth. Annotation quality depends on object detection, segmentation accuracy, and the subsequent image manipulation process.For copy-move forgery, the destination mask and source mask represent different regions. The destination mask identifies the pasted content, whereas the source mask identifies the original region from which the content was copied.For image splicing, the manipulation mask identifies the inserted region in the resulting composite image. Resizing, blending, geometric transformations, and other post-processing operations may cause differences between an object segmentation mask and the precise pixels affected by the manipulation.The masks should therefore be interpreted according to their documented annotation conventions. The dataset is intended for synthetic image-forensics research and model development; performance on real-world forensic cases requires independent evaluation on external datasets.5. Recommended UsesThe dataset supports the following applications:Image authenticity classification: Classifying images as authentic, copy-move forged, or spliced.Copy-move forgery detection: Detecting duplicated and repositioned content within an image.Image splicing detection: Identifying content inserted from a separate image.Manipulation localization: Predicting the spatial location of manipulated regions.Pixel-level segmentation: Training and evaluating segmentation models using manipulation masks.Deep learning benchmarking: Comparing CNN-based and other deep learning architectures for image forensics.Explainable AI: Investigating model predictions using techniques such as Grad-CAM, where appropriate.Robustness evaluation: Studying the effect of compression, resizing, blending, and other image-processing operations on forgery detection.For reliable model evaluation, training, validation, and testing partitions should be designed to minimize data leakage. Images derived from the same original source should ideally remain within the same partition.6. Data OrganizationThe dataset follows this directory structure: forensic_dataset/ ├── authentic/ │ └── images/ ├── copy_move/ │ ├── images/ │ ├── masks/ │ └── source_masks/ ├── metadata.csv ├── README.txt └── splicing/ ├── images/ └── masks/ The directories have the following purposes:authentic/images/: Contains authentic reference images.copy_move/images/: Contains images with copy-move manipulations.copy_move/masks/: Contains masks identifying the manipulated destination regions.copy_move/source_masks/: Contains masks identifying the original source regions used in copy-move operations.splicing/images/: Contains images with splicing manipulations.splicing/masks/: Contains masks identifying the inserted regions.metadata.csv: Stores dataset metadata and image-related information.README.txt: Contains dataset documentation and usage instructions.The dataset uses separate image and annotation directories for each forgery category. Consistent filenames or metadata references should be used to associate each image with its corresponding mask.
Source: Figshare
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7222491130
Cites:
Cited by:
Related to: