New AI Model Catches Tiny Document Forgeries by Fusing Seven Image Domains

Document forgery has entered an era in which a single swapped digit on a bank statement or a subtly altered stamp on an official certificate can slip past even a trained eye. As high-quality image editing tools proliferate across finance, government, judicial review, and archives management, the telltale signs of manipulation—broken character strokes, inconsistent noise, faint texture disorder—are becoming weaker and harder to isolate. A new study published in Discover Artificial Intelligence by Wenqi Zhao, Zhenjiang Li, and Lixin Wang of Gansu University of Political Science and Law tackles precisely this problem, presenting a detection framework that can locate tampered regions in document images even when the manipulated area occupies as little as one percent of the page.

The core insight behind the new method, called MDAF-FGE, is that no single way of looking at an image reveals all forgeries. Traditional forensic approaches relied on pixel statistics, edge operators, or template matching, but modern tampering destroys evidence across multiple physical layers of an image simultaneously: the JPEG compression pattern, the statistical structure of sensor noise, directional texture, and edge continuity. Deep learning methods, meanwhile, tend to prioritize semantic understanding—recognizing what a document says—over artifact perception, which means subtle manipulation traces are often washed out during deep feature abstraction. When a tampered region is tiny or the image has been heavily compressed, the forgery signal can vanish entirely before the network ever registers it.

To overcome this, the researchers built a multi-transformation domain generation module that decomposes each input document image into seven complementary feature domains. Four of these are extracted in parallel from the raw image: the ordinary spatial domain, the Discrete Cosine Transform domain that exposes compression inconsistencies, the Bayar Noise Domain that reveals residual noise patterns left by editing operations, and the Gabor Filter Response domain that captures directional texture anomalies. Three further domains are then derived hierarchically from these intermediate features: the Local Binary Pattern domain for local texture stability, the Roberts Cross-Gradient domain for edge discontinuities, and the Error Level Analysis domain, which highlights recompression artifacts that appear when a manipulated region has been saved at a different quality level than its surroundings.

Simply stacking seven feature sets, however, creates its own problem: redundant or noisy domains can drown out the informative ones. The multi-transformation domain attention fusion (MDAF) module addresses this with a three-part fusion network. First, domain-level channel attention assigns each domain an importance weight, learned automatically through global average pooling and a learnable weighting function, so the model can emphasize the domains that actually carry forgery evidence for a given image. Second, inter-domain correlation modeling captures the relationships between different domains, ensuring that complementary clues reinforce rather than compete with one another. Third, a dilated convolution spatial attention module applies multiple dilation rates to the fused features, capturing both fine local edge changes and broader structural inconsistencies across the document.

The fused features are then processed by a SwinUNet-V2 backbone, a hierarchical encoder-decoder built on Swin Transformer V2 blocks. The encoder passes the input through four stages in which feature dimensions grow from 96 to 768 channels while spatial resolution is progressively reduced through patch merging, allowing the network to model both fine-grained artifact details and long-range contextual dependencies through window-based self-attention. The decoder then restores spatial resolution through symmetric up-sampling and skip connections, preserving the fine spatial detail needed for precise localization, before a 1×1 convolutional prediction head with Sigmoid activation produces a pixel-level probability map of tampered regions.

Because document tampering regions are typically small and their artifacts weak, the authors added a second innovation: a fine-grained feature enhancement (FGE) module with three specialized branches. The edge enhancement branch amplifies gradient responses and boundary discontinuities associated with manipulated regions. The noise consistency branch performs differential learning to detect local statistical inconsistencies between authentic and altered areas. The multi-scale atrous convolution branch extracts spatial artifact responses at different receptive-field scales, catching both minute local traces and larger structural anomalies. The three branch outputs are concatenated, refined through convolution, batch normalization, and ReLU activation, and added back to the fused features with directional gradient terms that strengthen horizontal and vertical edge evidence.

Training the model relies on a joint multi-task loss that operates at three levels simultaneously. A binary cross-entropy loss constrains the classification accuracy of every individual pixel, while a region-level Dice loss optimizes the overlap between predicted and true tampered areas, keeping the model sensitive to small targets that a pixel-only objective would tend to ignore in favor of the dominant authentic background. A domain classification loss, weighted by a tunable coefficient, further constrains feature learning so the network does not over-rely on any single transformation domain, maintaining sensitivity to artifact patterns across all seven. The authors argue this decoupled design—separating global domain-level feature selection from local artifact enhancement—is what allows the system to handle both cross-domain coordination and weak-trace amplification, something simple feature concatenation cannot achieve.

The experiments were conducted on the CASIA Document Tampering Dataset released by the Institute of Automation of the Chinese Academy of Sciences, comprising 12,480 document images evenly split between 6,240 authentic and 6,240 tampered samples, with pixel-level ground-truth masks covering text replacement, copy-move manipulation, seal modification, and region deletion. Roughly 70 percent of images served for training and 30 percent for testing, with random cropping and flipping used for augmentation. The platform combined an Intel Core i9-13900K processor, 64 GB of DDR5 memory, and an NVIDIA GeForce RTX 4090 graphics card running Ubuntu 22.04 and Python 3.10, with image processing handled through OpenCV, NumPy, and scikit-image.

The results are striking, particularly under adverse conditions. Under severe JPEG compression with a quality factor of 10, MDAF-FGE achieved a mean intersection-over-union of 62.4 percent, well ahead of the multi-domain fusion baseline MFFD at 52.1 percent and the residual-based GFR-CNN at 43.6 percent; at quality factor 90 the figure rose to 85.7 percent. When tampered areas covered only one percent of the image, the method still delivered 79.3 percent precision and 63.7 percent recall, compared with 72.6 and 56.8 percent for MFFD and 61.4 and 47.5 percent for GFR-CNN. Against three recent detection frameworks—TALIU, DDT-Net, and a Hybrid CNN-Transformer model—the proposed method achieved the best score on all six evaluation metrics, including 93.4 percent precision, 91.2 percent recall, an F1 score of 92.3 percent, an AUC of 95.8 percent, and a Matthews correlation coefficient of 86.9 percent. Ablation experiments confirmed that removing the multi-domain input, the MDAF module, or the FGE module each degraded performance, with the full configuration reaching 91.3 percent balanced accuracy and a lowest balanced error rate of roughly 10.0 percent across four tampering types. Under noise interference, pixel accuracy stayed above 96 percent, and even at high noise levels the model retained about 93.8 percent accuracy when data augmentation was enabled.

The implications extend well beyond the laboratory. Banks verifying scanned statements, courts assessing digital evidence, and archives authenticating historical records all face documents that have been compressed, rescaled, and re-saved multiple times—the exact conditions under which older detectors fail. The authors are candid about remaining limitations, noting that generalization in extreme noise or long-text scenarios still has room for improvement, and that a broader unified comparison against more recent public-code methods under identical protocols is planned for future work. They point toward self-supervised artifact discovery, dynamic domain selection, and cross-modal document forensics as the next frontiers. For now, MDAF-FGE demonstrates that the future of document forensics may lie not in looking harder at a single image, but in looking intelligently at many versions of it at once—and letting attention decide which version tells the truth.

Subject of Research: Fine-grained detection of tampered regions in document images using multi-transformation domain attention fusion and fine-grained artifact enhancement

Article Title: Fine-grained tampered area detection method in document images based on multi-transformation domain attention fusion

Article References: Zhao, W., Li, Z., & Wang, L. (2026). Fine-grained tampered area detection method in document images based on multi-transformation domain attention fusion. Discover Artificial Intelligence, 6(1), Article 1323. https://doi.org/10.1007/s44163-026-02311-y

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02311-y

Keywords: document forensics, image tampering detection, attention mechanism, multi-domain fusion, deep learning, JPEG compression artifacts, Error Level Analysis, Swin Transformer, CASIA dataset, fine-grained artifact enhancement, multi-task learning, image forensics