Steganography hides a secret message in an ordinary file by making tiny changes, typically nudging a few pixel values or JPEG frequency coefficients by ±1. Steganalysis is the task of detecting those changes. At low payloads, below roughly 0.1 bits per pixel, the changes are buried in camera noise and texture, and even strong detectors are only a few points above chance. When those detectors fire, they cannot say where the hidden message is. A forensic analyst gets a probability from an opaque network and, at best, a saliency heatmap that may or may not reflect what the network actually used.
We propose a detector whose final score is, by construction, a sparse sum of evidence from individual image blocks. Each block's evidence comes from two views of the same JPEG file, the decompressed pixels and the quantized frequency coefficients. Each view's features are split into a part shared across views, which is mostly image content, and a part private to that view, where we hypothesize the weak stego trace survives. The shared content features decide which blocks to read and how to weight the two views. Only the private features supply evidence. Steganalysis has an unusual advantage for testing this: we generate the stego images ourselves, so the true locations of the changes are known. We can therefore measure whether the model's explanation is correct, not just whether it looks plausible.
Core Claim
We make two claims. First, at low payloads, a block-sparse additive readout over view-private evidence can match the accuracy of an equal-capacity black-box head, while its block attributions recover the ground-truth change locations far better than post-hoc saliency maps. Second, this depends on the decoupling. If the per-block evidence is computed from features that mix content and residue, or from features pretrained to align the two views, both detection and attribution degrade. The reason is that cross-view alignment keeps what the two views agree on, which is the scene, and discards the faint view-specific statistics the payload leaves behind.
The second claim is a hypothesis, not an assumption. A JPEG-domain change also appears in the pixels, so the stego trace is not strictly private to one view. What we claim is that its statistical form differs between views, for example breaks in dependencies across 8×8 block boundaries versus distortions in coefficient histograms, and that an objective rewarding cross-view agreement throws this away. We test this directly by probing which subspace carries the signal.
Where This Sits
Low-rate detectors have improved through fixed noise-extracting filters and better aggregation. CIS-Net suppresses cover texture with fixed truncation and sub-linear pooling, and visualizes decisions with post-hoc class activation maps. CTS-Net adds global self-attention to spatial noise residuals. Zeng et al.'s hybrid JPEG detector runs 75 parallel subnets over frequency-domain residuals and concatenates their outputs. All three end in opaque classification heads, and none separates content from residue across views. They motivate our proposal, because they show the trace exists and that fixed suppression helps. They also constrain it: Zeng et al. show that fixed quantization and truncation of residuals is worth roughly 20 accuracy points on J-UNIWARD, so our view encoders keep those front ends rather than learning them from scratch.
The decoupling technique comes from DeCUR. DeCUR pushes the cross-view correlation of "common" dimensions toward the identity and that of "unique" dimensions toward zero. It found that the shared space between very different sensors is sparse, and that preserving unique dimensions improves robustness. We reuse its loss, and we differ in three ways. Our signal of interest lives in the private part rather than being a side benefit. Our common/private split is not used statically but drives a per-block gate. And we test on microscopic perturbations rather than scene semantics.
The readout comes from NGSLL. NGSLL has a deep network generate sample-specific sparse linear weights over a simple representation, selected by a differentiable top-K gate, so the prediction is exactly a linear function of that representation. It matched black-box accuracy on binary MNIST, but it has only been tested on high-signal tasks with a single view. We change the simple representation from pixels or words to per-block, per-view private evidence.
We take seriously the warning that transparency is necessary but insufficient for comprehensibility. Our explanation is deliberately coarse: which blocks, and which view. That is the unit an analyst can check against an extracted change map.
Research Program
Each JPEG image is cut into 8×8 blocks aligned with the compression grid. Two encoders run over the image. The pixel encoder applies fixed noise-extracting high-pass filters (SRM-style) to the decompressed image. The frequency encoder applies Zeng-style filters with fixed quantization and truncation to the coefficient planes. Each encoder emits, per block $b$, a common vector $c_b$ and a private vector $u_b$. A DeCUR-style loss on cover images aligns $c^{\text{pix}}_b$ with $c^{\text{dct}}_b$ and decorrelates $u^{\text{pix}}_b$ from $u^{\text{dct}}_b$. A small linear map turns each private vector into a scalar evidence value, $e^{\text{pix}}_b$ and $e^{\text{dct}}_b$.
A gating network reads only the common features of all blocks, with global context, and produces two quantities per block. The first is a selection weight $g_b(x)$, passed through an NGSLL top-K gate with $K$ annealed from large to small. The second is a view-mixing weight $\alpha_b(x) \in [0,1]$. The detection score for image $x$ is
$$s(x) = \sum_{b} g_b(x)\,\big[\alpha_b(x)\, e^{\text{pix}}_b + (1-\alpha_b(x))\, e^{\text{dct}}_b\big],$$
where at most $K$ of the $g_b$ are nonzero. The score is exactly the sum of the displayed per-block terms, so each term is a faithful attribution.
Routing on content is the adaptive-fusion component, and it is motivated by how modern steganography works. Content-adaptive embedders such as WOW, S-UNIWARD and J-UNIWARD place changes in textured regions. Content therefore predicts where to look, while residue determines whether anything is there. This also replaces DeCUR's fixed global common-to-unique ratio with a per-block, per-image allocation.
Training uses binary cross-entropy on $s(x)$ plus the decoupling loss. Following CIS-Net, we use a payload curriculum from 0.4 down to 0.1 and 0.05.
Baselines
The naive reference point is the evidence of the ALASKA challenge. There, the winning entry reached a miss-detection rate of around 55% at 5% false alarms, and most entries sat between 55% and 58%. Low-rate, realistic detection is close to the floor, and any method should be reported relative to chance.
The strongest existing spatial-domain method is CTS-Net. It reports 75.42% detection accuracy on BOSSbase 1.01 for WOW at 0.1 bpp, against 70.65% for SRNet in the same paper, and 69.50% on 1024×1024 ALASKA#2 images. It is the strongest competitor because it is the best reported low-rate number in our file and is already parameter-efficient. We retrain it with our splits, curriculum and tuning budget rather than quoting it. Beating it means matching its accuracy within one point at 0.1 bpp while additionally delivering faithful attributions. Exceeding it by more than seed variance would be a bonus, not the claim.
For JPEG, the strongest file-reported baseline is Zeng et al.'s hybrid detector at 74.5% on J-UNIWARD (ImageNet) with fixed quantization and truncation. The file does not state the payload for this number, so we re-run it at matched payloads. For CIS-Net we found no reported 0.1 bpp number on Lacuna; its 14.62% error is for S-UNIWARD at 0.4 bpp.
Three ablations isolate our ingredients, all using the same encoders, data and compute.
- Entangled evidence. Keep the additive block head but compute evidence from unsplit features. This tests the decoupling.
- Static fusion. Fix $\alpha_b=1/2$ and use uniform selection. This tests the content gate.
- Black-box head. Put an MLP on the same decoupled features. This tests what the additive constraint costs.
A fourth variant pretrains both encoders with a plain cross-view contrastive objective before fine-tuning. It is the direct test of the over-alignment hypothesis.
Every model gets the same number of hyperparameter trials, three seeds, and identical cover/stego pair splits. Splits are made at the cover level, so no cover appears in both train and test.
Why not just…?
Why not just train CTS-Net and explain it with Grad-CAM or SHAP? Post-hoc maps approximate a black box from outside. They tend to highlight textured regions, which is exactly where content-adaptive embedders put changes, so they can look right while ignoring the actual changes. SHAP-style perturbations also sample off-manifold inputs (Axiomatic On-Manifold Shapley). We settle this empirically by scoring both methods against the ground-truth change maps with the same metrics. To make the comparison fair, we include a texture-only predictor as a control: if Grad-CAM does no better than a texture map, it is not explaining stego.
Why not just concatenate pixel and frequency streams into one strong detector with attention pooling over blocks? This is the real threat. Attention-pooled, multiple-instance-style detectors are already additive over blocks and therefore partly faithful. Ablation 1 is exactly this model with our head, so the comparison is clean. If entangled evidence matches decoupled evidence on both accuracy and attribution precision at 0.1 and 0.05 bpp, the decoupling claim fails. We would then report the simpler model as the result.
Why not just use a single pixel-domain detector, since JPEG changes show up in pixels anyway? Zeng et al.'s 20-point gain from frequency-domain quantization and truncation suggests that the frequency view carries a form of the signal the pixel view does not expose easily. The learned $\alpha_b$ distribution and a pixel-only variant of our model test this directly. If the gate collapses to one view, the two-view design is unnecessary.
Proposed Experiments and Decisive Checks
We use four settings: BOSSbase 1.01 with WOW, S-UNIWARD and HILL at 0.1 and 0.05 bpp; BOSSbase compressed to JPEG with J-UNIWARD and UERD at 0.1 and 0.05 bits per nonzero AC coefficient; an ImageNet JPEG subset following Zeng et al.; and ALASKA#2 pairs, whose change maps come from diffing cover and stego files. We report accuracy, and miss detection at 5% false alarm for comparability with ALASKA.
The decisive attribution test is a causal, on-manifold intervention. For each detected stego image, we replace the top-K attributed blocks with the corresponding real cover blocks and measure the score drop. We compare this to replacing K random textured blocks, and to replacing the top-K blocks chosen by Grad-CAM on CTS-Net. Because the substituted blocks are genuine covers, the intervention stays on the data manifold. We also report the fraction of true changes falling inside the selected blocks, as a function of K.
The over-alignment check uses linear probes for stego presence on $c$ versus $u$, and on contrastively pretrained features. We expect the stego signal to concentrate in $u$ and to be degraded in the contrastive features, with the gap widening as payload falls. Cover-source robustness is tested by training on one ALASKA processing pipeline and testing on another, following SimMMDG's argument that private evidence is more robust to shift.
The proposal fails if any of the following hold at 0.1 bpp:
- the additive head loses more than about two points to the black-box head on the same features;
- cover-block substitution in the selected blocks does not drop the score more than substitution in random textured blocks;
- entangled evidence performs as well as decoupled evidence.
References
- CIS-Net: A Novel CNN Model for Spatial Image Steganalysis via Cover Image Suppression
- Image steganalysis based on CNN-Transformer
- Large-Scale JPEG Image Steganalysis Using Hybrid Deep-Learning Framework
- Decoupling Common and Unique Representations for Multimodal Self-supervised Learning
- Neural generators of sparse local linear models for achieving both accuracy and interpretability
- (Un)reasonable Allure of Ante-hoc Interpretability for High-stakes Domains: Transparency Is Necessary but Insufficient for Comprehensibility
- The ALASKA Steganalysis Challenge: A First Step Towards Steganalysis
- Axiomatic On-Manifold Shapley via Optimal Generative Flows
- SimMMDG: A Simple and Effective Framework for Multi-modal Domain Generalization