Distilling Trace Survival: Reference-Trained Reliability Maps for Laundering-Robust Forgery Localization

Forgery localizers, which mark the edited pixels in a photo, usually combine two views of the image. One view reads what the image shows: objects, lighting, edges. The other reads faint low-level traces left by the camera and by JPEG compression. The second view catches edits that look seamless. It also breaks first when the image is "laundered," meaning recompressed, resized or blurred, as social media platforms do routinely. When that happens, the noise view does not simply go quiet. It sends corrupted evidence into the fusion step, and the detector stays confident anyway. We propose to give the localizer a blind, per-pixel estimate of how much of each forensic trace survived laundering. The estimate is learned from a teacher that can only be computed during training, because only then do we hold the image before laundering and can measure, region by region, how far each trace drifted. The fusion step then uses this map to lean on the noise view where its traces survived and on the semantic view where they did not.

The idea borrows from full-reference metric distillation. In that line of work, a quality score computed against a clean original serves as a free training label for a model that never sees the original at test time. It also borrows from multi-stream forensic fusion. What is new is the teacher signal. We do not distill perceptual similarity such as SSIM, and we do not distill a single global confidence number. We distill a dense map of forensic trace survival for each stream and use it inside fusion, not as a filter applied after the fact.

Core Claim

Supervising a fusion module's per-stream reliability maps with full-reference trace-survival targets should improve localization under laundering, especially laundering that never appeared in training. The comparison is against three alternatives: the same architecture whose confidence is learned only from the localization loss, the same architecture supervised with perceptual full-reference maps, and the same architecture supervised with global laundering parameters. The claim fails if any of these alternatives matches the distilled model on laundered and unseen-laundering benchmarks.

Where This Sits in Prior Work

The closest work is Enabling Your Forensic Detector Know How Well It Performs on Distorted Samples (DACOM). It uses full-reference image quality as training-time privileged information and finds that detector accuracy falls roughly linearly with that quality. It then trains a blind regressor that predicts a global detectability score for a frozen synthetic-image classifier, reporting 97.73% linear correlation with actual accuracy on ForenSynths. DACOM motivates this proposal in two ways. It shows that reference-derived targets predict forensic failure, and it reports that perceptual no-reference metrics predict forensic accuracy poorly. That second finding is why we move the teacher from perceptual quality to trace survival. We also differ from DACOM in three further respects: our output is dense rather than scalar, it is specific to each stream, and it is trained jointly with the localizer rather than added post hoc.

The method for distilling dense maps comes from Blind Predicting Similar Quality Map for Image Quality Assessment. That paper trains a U-Net to predict FSIM- or SSIM-style similarity maps without the reference image and reaches 0.862 SRCC on TID2013. We reuse its target-map regression but change what the map measures. Two papers supply evidence that spatial quality maps help fusion. Quality-Aware Selective Fusion Network for V-D-T Salient Object Detection suppresses low-quality regions of visible, depth and thermal features and reports a 16.00% MAE improvement over MFFNet. A Quality-Guided Mixture of Score-Fusion Experts Framework for Human Recognition shifts biometric fusion weight away from degraded modalities. In both papers the quality labels come from agreement with the task's ground truth or from ranking, not from a physical reference measurement, and neither is evaluated in forensics.

On the forensic side, three systems serve as both host architectures and constraints. TruFor fuses RGB features with its learned camera-noise fingerprint, Noiseprint++, through a cross-modal transformer. It already outputs a per-pixel confidence map, but that map is learned only from the localization loss, and the model's robustness comes from training on more than 500 laundering histories. CAT-Net adds a DCT stream, and its p-F1 on GRIP falls from 76.45% to 24.60% without double-JPEG pretraining, which shows how brittle compression traces are. CO-SPY reweights semantic and artifact features with a learned regulator that has no explicit degradation target. Our proposal is effectively TruFor's confidence idea, made per-stream and given a measured target.

The Teacher Signal

Training forgeries are laundered synthetically, so each training sample comes with its pre-laundering version $x$ and its laundered version $\tilde{x}$. If laundering included resizing, $\tilde{x}$ is resampled back onto $x$'s pixel grid. For each low-level stream $k$ we use a frozen trace extractor $\phi_k$: Noiseprint++ residuals for the noise stream and blockwise DCT coefficient statistics for the compression stream. The survival target at pixel $p$ is the local normalized correlation between the clean and laundered traces:

$$ s_k(p) = \mathrm{corr}_{W(p)}\big(\phi_k(x),\, \phi_k(\tilde{x})\big), $$

Here $W(p)$ is a window centered on $p$, 16×16 for noise and the enclosing 8×8 block grid for DCT. The value is clipped to $[0,1]$.

This target varies across the image for two reasons. Flat or saturated regions lose noise traces faster than textured ones under the same compression, and the edited regions themselves have different trace statistics. A global laundering parameter cannot express either effect.

We also compute a second, task-grounded target as a variant. It is the per-pixel drop in a single-stream localizer's accuracy between $x$ and $\tilde{x}$. This lets us test whether trace-level or detectability-level supervision is the better teacher.

A blind head $g_k$ shares the localizer's encoder and predicts $\hat{s}_k$ from $\tilde{x}$ alone. Training minimizes

$$ \mathcal{L} = \mathcal{L}_{\text{loc}} + \lambda \sum_k \big\| \hat{s}_k(\tilde{x}) - s_k \big\|_1 , $$

where $\mathcal{L}_{\text{loc}}$ is the usual localization and detection loss and $\lambda$ weights the distillation term. Inside the fusion step, $\hat{s}_k$ scales stream $k$'s keys and values before cross-modal attention, and it replaces TruFor's learned confidence when the final integrity score is pooled.

Research Program

We build on a re-implementation of TruFor's RGB plus Noiseprint++ backbone, add a DCT stream in the style of CAT-Net, and keep all trace extractors frozen. Every training forgery is passed through a laundering pool covering JPEG at varied quality factors, resizing, blur, noise, and chained combinations. Some laundering families are held out entirely for testing: WebP recompression, real platform recoding as in DSO-1 through Facebook, and unseen chain orders.

The work answers three questions in order. First, can a blind model predict $s_k$ at all? We measure per-pixel correlation between predicted and true survival maps on held-out laundering. Second, does gating fusion with $\hat{s}_k$ improve localization over the strongest baselines? Third, does the benefit come from the trace-survival target specifically? This is the central ablation.

Baselines

The naive reference point is the dual-stream localizer with plain concatenation fusion and no reliability map, trained with the same laundering pool. It measures what fusion alone achieves without any confidence mechanism.

The strongest existing method is TruFor. It reports an average localization F1 of 0.785 and detection AUC of 0.857 across seven benchmarks, and F1 of 0.685 on DSO-1 under Facebook compression against 0.484 for the next competitor (TruFor). It is the right competitor because it already has a per-pixel confidence map and laundering-hardened noise traces. Any gain we report must therefore come from the supervision target, not from having a confidence map at all. We compare against the published checkpoint and also against a TruFor retrained on our exact data, laundering pool, number of training steps and hyperparameter search budget. Our method must use no extra capacity beyond the small survival head, which we match by giving the retrained TruFor an equally sized auxiliary head. For localization specifically under compression, CAT-Net's 76.45% p-F1 on GRIP (CAT-Net) is the reference figure. We count the proposal as beating TruFor if it improves F1 on DSO-1 Facebook and on held-out laundering families beyond seed variance, while staying within noise of TruFor's clean-benchmark average.

Three ablations isolate the key ingredient, and each keeps the architecture fixed while changing only the target. The first supervises the same head with SSIM or FSIM maps computed between $x$ and $\tilde{x}$, the perceptual teacher of BPSQM. The second supervises with global laundering parameters broadcast to every pixel. The third applies a DACOM-style global detectability scalar after fusion (DACOM), using its quality and distortion descriptors. DACOM itself was evaluated on whole-image synthesis detection and has no reported localization result. No reported result was found on Lacuna for DACOM on DSO-1, CASIA or CocoGlide, so we report it only as our own re-implementation.

Why not just…?

Why not just trust TruFor's end-to-end confidence map? This is the most serious threat. A confidence map trained only on localization loss could already learn where traces died. Our evidence is suggestive rather than decisive: TruFor still loses about 0.1 F1 under Facebook compression despite its augmentation. The matched retraining with an equal-size auxiliary head settles the question directly. If that baseline ties us on held-out laundering, the distillation adds nothing and we will report it as such.

Why not just distill perceptual full-reference maps like SSIM? DACOM reports that perceptual quality predicts forensic accuracy poorly, since mild recompression can look clean while erasing noise traces. The perceptual-target ablation tests this in the dense setting. We expect the gap to be largest on high-quality JPEG and resizing, where perceptual similarity stays high but trace survival collapses.

Why not just predict the laundering parameters, which are free labels? This is simpler and needs no reference computation, so it is a real competitor. Its weakness is that one quality factor yields very different survival in flat and textured regions, and real platform chains have no single parameter. The broadcast-parameter ablation, evaluated on DSO-1 Facebook and WebP, which are outside the parameter vocabulary, will show whether spatial variation matters.

Why not just augment harder? CAT-Net's GRIP collapse without double-JPEG pretraining and TruFor's residual drop show that augmentation hardens traces but does not tell fusion when they failed. Every baseline here shares our laundering pool, so augmentation is held constant and cannot explain any difference.

Proposed Experiments and Decisive Checks

We train on CASIAv2, Fantastic Reality and synthetically laundered copies of them. We evaluate on CASIAv1, Columbia, COVERAGE, NIST16, IMD2020, CocoGlide, GRIP and DSO-1, with Facebook-compressed DSO-1 as the key real-laundering test. Each test set also gets a controlled severity sweep under both seen and held-out laundering families. The metrics are pixel F1 and IoU for localization, AUC for detection, and an abstention curve that measures localization F1 on the top fraction of images ranked by mean predicted survival.

The decisive check combines two results. The trace-survival model must beat the matched TruFor retrain and the perceptual and parameter ablations on held-out laundering. Its predicted survival maps must also correlate with true survival on unseen laundering. If the maps are accurate but localization does not improve, the bottleneck is fusion, not reliability estimation. If localization improves under the parameter target as much as under ours, the spatial teacher is unnecessary. Either outcome would redirect the work, and both are informative.