Every camera leaves a faint, pixel-level noise pattern in its images, caused by tiny manufacturing differences between sensor elements. When someone pastes in a region from another image, that pattern breaks inside the pasted region, and forensic detectors look for the break. The break is weak, so a detector has to pool evidence over many pixels before it can be confident. A forgery covering 0.3% of the image offers few pixels to pool. Standard convolutional backbones then make things worse: pooling and repeated convolutions smooth away exactly this low-level signal before the layers that make the final per-pixel decision. We propose to measure that loss directly and to fix it. The fix is a full-resolution noise stream built from residual dense blocks, so every deeper layer receives the untouched noise residual alongside the scene context it has built up. We test it on a benchmark where forgery size and compression strength are controlled.
The claim is narrow and falsifiable. For blind, single-image localization (no reference fingerprint for the camera), the smallest forgery a learned detector can find is set mainly by how much of the noise residual survives to its decision layers, not by how the loss handles class imbalance. If so, dense reuse of the residual should shift the detection-versus-area curve toward smaller forgeries. The gain should be largest for forgeries under 1% of the image, near zero for large ones, and should vanish for copy-move forgeries and heavy JPEG compression. In those two cases the noise contrast itself is absent, and no architecture can recover it.
Why this framing
The physics gives a simple expectation. Suppose the forged region shifts the per-pixel noise statistic by $\delta$ relative to a per-pixel spread $\sigma$, and a detector pools $n$ roughly independent residual pixels. The test statistic then grows like $z \approx \delta\sqrt{n}/\sigma$, the standard result for averaging independent noisy measurements. Two consequences follow. Detectability is governed by forgery area, since $n$ scales with area. Anything that attenuates the residual before the decision layer lowers the effective $\delta/\sigma$, which raises the minimum detectable area quadratically.
The classical, non-blind literature already shows this trade-off. Guided filtering for PRNU-based localization of small-size image forgeries correlates against a known camera fingerprint, called PRNU (photo-response non-uniformity). It shows that large averaging windows bias decisions on small forgeries, and that content-adaptive pooling with a guided filter raises AUC on 48×48 forgeries from 0.63 to 0.78. That paper motivates our size-stratified view. We do not assume its reference fingerprint is available.
Deep forensic models extract the residual blind, at the input. Learning Rich Features for Image Manipulation Detection uses fixed steganalysis filters. Learning Discriminative Noise Guidance for Image Forgery Detection and Localization (NGNet) subtracts the output of a learned denoiser, applies a constrained convolution, and adds a divergence loss. Image Forgery Localization via Guided Noise and Multi-Scale Feature Aggregation uses guided and Sobel filters followed by a dilated pyramid. All three then hand the residual to a ResNet, an EfficientNet, or a cross-attention stage, where it gets mixed and downsampled.
TransForensics: Image Forgery Localization with Dense Self-Attention reports that its shallow feature levels contribute more than its deep ones. That is indirect evidence that forensic signal is lost with depth. TransForensics has no explicit noise stream and no mechanism to carry early features forward intact. Our proposal supplies that mechanism and tests it on precisely the size regime where the loss should matter.
Core design
The method comes from Residual Dense Network for Image Super-Resolution (RDN). In RDN's residual dense blocks, each convolution receives the concatenation of all earlier outputs in the block, $x_l = H_l([x_0,\dots,x_{l-1}])$. A 1×1 convolution then compresses the stack, and an identity skip adds the block input back. Super-resolution and restoration networks such as MFRNet use this design because it keeps fine high-frequency detail available deep into the network. That is the property a fragile noise residual needs.
We build a noise stream that never downsamples. It takes the denoiser residual (CBDNet followed by a constrained convolution, as in NGNet) and passes it through a stack of residual dense blocks with a narrow growth rate of 8 to 16 channels. A separate RGB backbone builds semantic context at the usual reduced resolutions. At each backbone scale, the context is upsampled and fused with the current noise-stream state. The raw residual $x_0$ is also re-concatenated into every block. As a result, the per-pixel decision head always sees the unattenuated residual together with context that can tell genuine edges from seams.
The known cost of dense connectivity at full resolution is memory blow-up. A noise residual is low-dimensional, so a narrow stream makes this tractable. We report memory and latency rather than assume it.
Imbalance handling is a controlled factor, not a headline ingredient. Training on crops means many patches contain no forged pixels. Under that condition, Gradient-Aware for Class-Imbalanced Semi-supervised Medical Image Segmentation shows that standard Dice loss gives no gradient for the absent class. Its replacement losses, GADice (Dice with a gradient that persists when the target is absent) and GACE (cross-entropy restricted to the hardest pixels, with a balancing weight), took two small abdominal organs from 0% to 27–51% Dice on Synapse CT scans. We cross the architecture with three losses: Dice+Focal as in TransForensics, GADice+GACE, and plain cross-entropy. This separates gains due to evidence preservation from gains due to optimization.
Research Program
The first deliverable is a controlled micro-forgery benchmark. We splice regions between images from different cameras into hosts at 768×1024 resolution. Forgery areas are swept log-uniformly from about 0.05% to 10% of the image. Square patches from 48×48 upward reproduce the protocol of the guided-filter PRNU paper (48×48 is about 0.29% of such an image), and irregular object masks are included as well. Each image is re-encoded at JPEG qualities 95, 75, 50 and 30. Using the same pipeline, we create copy-move forgeries, which serve as a negative control because the region shares the host's sensor.
The second deliverable is a diagnostic. We train linear probes on frozen features at each depth of every architecture to predict the per-pixel forgery label. Probe accuracy as a function of depth and forgery area gives a direct "residual survival" curve. This tests the dilution hypothesis without relying on end-task scores.
The third deliverable is the architecture and the factorial comparison described below. We also evaluate on the public benchmarks CASIA, COVERAGE, NIST16, Columbia and IMD2020, stratifying every result by forgery area.
Baselines
The naive reference is a single-stream RGB segmentation network on the same backbone with no noise input. It shows how much any noise branch contributes at each forgery size.
The strongest existing method on our target regime is the guided-noise model of Image Forgery Localization via Guided Noise and Multi-Scale Feature Aggregation. It reports 87.1% AUC on forgeries under 1% of the image, against PSCC-Net at 82.9%, and an F1 gain of 19.3% over HiFi-Net on NIST16. NGNet is the strongest noise-guided method overall, reporting 99.81% AUC and 98.72% F1 on CASIA-D and 94.1% AUC on COVERAGE. TransForensics represents the shallow-features-plus-imbalance-loss approach, reporting AUC 0.850 and F1 0.627 on CASIA and AUC 0.884 on COVERAGE. RGB-N reports AUC 0.937 on NIST16.
All of these are retrained on our training split with the same input resolution, crop policy and augmentation. Each gets the same hyperparameter search budget (an equal number of trials), and models are matched within 10% in parameters and FLOPs. Where we only quote published numbers, we say so.
Beating the strongest baseline means a statistically significant gain in pixel AUC and F1 on the stratum under 1% area, averaged over at least three seeds, at uncompressed and Q=75. It also means a lower minimum detectable area, defined as the smallest area at which pixel AUC exceeds 0.8. Parity at large areas is expected and acceptable.
The non-blind guided-filter PRNU method is run on the subset where reference fingerprints can be built. It reports 0.78 AUC on 48×48 and 0.85 on 64×64 forgeries. We treat it as an informed upper reference, not as a competitor.
The key ablations isolate dense reuse. The first replaces the residual dense blocks with plain residual blocks of matched width. The second keeps a full-resolution plain stream but feeds the raw residual only once, through a single long skip into the head. The third uses U-Net-style skips at every scale but without dense concatenation. The fourth substitutes a dilated pyramid at matched cost.
Why not just…?
Why not just add one long skip from the noise residual to the output head? This is the most serious threat. If dilution is the whole story, a single late skip might recover most of the gain. Dense reuse should help only if context and noise need to interact repeatedly at many depths, for instance to suppress residual spikes at natural edges. We do not know which is true. The single-skip and U-Net-skip ablations settle it on the under-1% stratum. If they match the dense stream, we will report that the cheap fix suffices and keep the diagnostic and benchmark as the contribution.
Why not just use Dice plus Focal loss, or better imbalance losses, on an existing detector? TransForensics already uses Dice+Focal and still produces coarse block-like intermediate maps. The gradient-aware losses fix a real failure, the missing gradient on patches without forged pixels, as their Synapse results show. However, a loss cannot restore evidence the network has already discarded. The factorial design measures the two effects separately. If GADice+GACE on a plain backbone closes the gap, our architectural claim fails.
Why not just use dilated convolutions, as the guided-noise model does? Its pyramid keeps resolution but samples the feature map at large gaps. It also degrades under JPEG Q≤50, which is consistent with noise being under-preserved. The matched-cost dilated ablation and the depth-wise probe curves will show whether the residual survives the pyramid as well as it survives dense blocks.
Why not just adapt a segmentation foundation model with small adapters? Detective SAM reaches 36.99 IoU on out-of-distribution generative edits with under a million adapter parameters. Its authors, however, state it is not designed for copy-move or classical splicing, and it depends on hand-set perturbation strength. We include it on the splicing benchmark as a reference rather than a primary baseline, and we expect it to trail on tiny spliced regions.
Decisive checks
Four outcomes would confirm the hypothesis together. First, probe curves show the residual decaying with depth in standard backbones but not in the dense stream. Second, the end-task gain over the matched plain stream and over the single-skip ablation is concentrated below 1% area and shrinks toward zero above 5%. Third, the minimum detectable area improves at Q=95 and Q=75 but not at Q=30. Fourth, the gain is absent on our copy-move control and small on COVERAGE.
NGNet's 94.1% COVERAGE AUC partly challenges the last prediction. It suggests that learned noise features also capture resampling or boundary cues, so a modest gain there would be read as boundary evidence rather than sensor evidence. If gains instead appear uniformly across sizes, or vanish under the single-skip ablation, the dense-reuse claim is rejected. Under the single-skip outcome, the contribution reduces to the benchmark, the diagnostic, and that simpler fix.
References
- Guided filtering for PRNU-based localization of small-size image forgeries
- Learning Rich Features for Image Manipulation Detection
- Learning Discriminative Noise Guidance for Image Forgery Detection and Localization
- Image Forgery Localization via Guided Noise and Multi-Scale Feature Aggregation
- TransForensics: Image Forgery Localization with Dense Self-Attention
- Residual Dense Network for Image Super-Resolution
- MFRNet: A New CNN Architecture for Post-Processing and In-loop Filtering
- Gradient-Aware for Class-Imbalanced Semi-supervised Medical Image Segmentation
- Detective SAM: Adaptive AI-Image Forgery Localization