Steganography hides messages inside images, and recent invertible neural networks — networks built to run equally well forward (embedding) and backward (extraction) — can now hide entire images inside one another at high payload density. But virtually all of this assumes a pristine digital channel: the container file travels untouched from sender to receiver. The moment a container is printed on a surface and re-captured by a camera, pixel-exact assumptions collapse, and the hidden bits die exactly the way imperceptible adversarial noise dies. Meanwhile, the adversarial-robustness literature has spent years mapping which image components survive lossy transforms and physical capture: high-frequency pixel configurations are destroyed by compression, filtering, and camera physics — which is why fast spectral defenses strip them Feature Distillation and why diffusion purification scrubs protective perturbations IMPRESS — while low-frequency macro structure survives, which is precisely how physical camouflage attacks persist in the real world DTA.
The proposal here is to use that fragility map as a design rule and build the first steganography system engineered for the print-to-camera channel: train an invertible embed/extract pair with a differentiable model of printing and photography inside the training loop, deliberately place the payload in the low-frequency band that physical transforms preserve, and distill the result into a feed-forward network so that both writing and reading a payload fit inside a single video frame. The application regime is cooperative physical authentication — a surface that carries verifiable identity or telemetry readable by an ordinary camera at frame rate — with an adversarial evaluation in which an attacker tries to scrub the mark after capturing it.
Core Claim
High-capacity covert payloads (at least 1 bit per pixel, or short identity strings at proportionally higher reliability) can survive the print-to-camera channel and be written and read in real time, if and only if two ingredients are combined: a differentiable transformation network modeling the physical channel during training, and payload placement in the macro-structural band that both real-time defenses and purification pipelines leave intact. Neither ingredient alone suffices, and no existing system has either.
Prior Work and the Gap
Four lines of work converge on this intersection without meeting it. Robust Invertible Image Steganography is the strongest robust high-capacity system: it hides up to five full secret images in a host with roughly 36 dB secret-reconstruction quality, and survives Gaussian noise ($\sigma=10$, at 28.08 dB on DIV2K where a standard invertible scheme collapses to 8.55 dB) and JPEG down to quality factor 40. But its channel model is purely synthetic digital noise; it has never been evaluated against camera capture, and no reported result found on Lacuna for any steganography scheme evaluated through a physical print-to-camera pipeline. Fixed Neural Network Steganography establishes the capacity ceiling — 0.0% bit error at payloads up to 3 bits per pixel — while itself documenting that its payloads are destroyed by lossy operations and that per-image embedding takes 10 seconds to 3 minutes, disqualifying it from any real-time pipeline. Semantic Deep Hiding supplies the key design principle I reuse: it showed that moving hidden content from additive pixel noise into structured, semantically meaningful patterns embedded via invertible networks and wavelets defeats eighteen countermeasures (holding CIFAR-10 test accuracy to 33.01% and ImageNet to 11.39% in its unlearnability setting) — but it was evaluated only against digital filters, never a physical channel, and it hides an image, not a decodable message under a bit-error metric. Finally, DTA and RAUCA built exactly the machinery I need for the channel — neural proxies of rendering, lighting, and weather, validated physically at 1:10 scale (DTA drops physical-scale EfficientDet AP from 0.94 to 0.34) — but used it only to optimize loud, conspicuous attack textures. DynamicPAE proves that a feed-forward generator trained against noisy physical gradients can produce scene-adaptive patterns at 12 ms per frame, but only for localized high-contrast patches, not covert data. InvisMark shares the intent of robust invisible provenance marking, but reports no physical-camera or latency evaluation found on Lacuna. The gap is specific: nobody has put a differentiable physical channel model inside an invertible steganography training loop, and nobody has distilled high-capacity embedding to frame rate.
Research Program
The system has three components. First, the channel: I will train a differentiable transformation network in the style of DTA's DTN, conditioned on viewpoint, distance, lighting, and weather, using the DCI benchmark simulation environment (seven scenes, 40 angles, 15 distances, over 20,000 positions) as ground truth for what print-and-photograph actually does to a surface. Second, the codec: an invertible embedding/extraction pair in the RIIS lineage, trained with the transformation network in the loop. Writing $x$ for the cover image, $m$ for the payload bitstring, $E$ and $D$ for the embedding and extraction halves, and $g_\phi(\cdot,\tau)$ for the transformation network applied under a randomly sampled physical condition $\tau$, training minimizes
$$\min_{E,D}\; \mathbb{E}_{x,\,m,\,\tau}\Big[\ell_{\mathrm{bit}}\big(D\,\circ\,g_\phi(E(x,m),\tau),\,m\big)\;+\;\lambda\,\ell_{\mathrm{perc}}\big(E(x,m),\,x\big)\Big]$$
where $\ell_{\mathrm{bit}}$ counts payload bit errors after the simulated physical channel and $\ell_{\mathrm{perc}}$ keeps the container visually indistinguishable from the cover. The fragility literature dictates a structural choice inside this objective: the invertible coupling layers are constrained to route payload information into the low-frequency wavelet subbands and macro-semantic components — the components that Feature Distillation deliberately preserves for accuracy and that physical camouflage DTA exploits for survival — rather than the high-frequency subbands where adversarial noise and fragile steganographic payloads both live. A soft-decision error-correcting code sits between $E$ and $D$ so the extractor outputs reliability-weighted bits. Third, the real-time path: iterative embedding is distilled into a feed-forward encoder using DynamicPAE's residual-driven trajectory guidance, which was invented precisely to stabilize training against noisy gradient feedback from non-differentiable physical simulation; the goal is constant-time embedding and extraction, each under 27 ms, matching the operational envelope that Themis (36–59 fps) and DynamicPAE (12 ms) establish for real perception systems.
Baselines
The naive reference point is RIIS run exactly as published — trained against its synthetic noise channel and then pushed through the physical pipeline. RIIS is the right naive baseline because it is the current best robust high-capacity scheme: it reports approximately 36 dB secret PSNR hiding up to five secret images, 28.08 dB under Gaussian noise $\sigma=10$ on DIV2K, and JPEG resilience down to quality factor 40, per Robust Invertible Image Steganography. Its documented behavior is that bijective schemes degrade sharply off-channel (the ISN baseline collapses to 8.55 dB under the same Gaussian noise), so I expect near-total payload loss through the camera channel; this baseline establishes that the gap is real rather than assumed. No reported result found on Lacuna for RIIS or any invertible steganography scheme under camera capture, which is itself the point.
The strongest competitor is RIIS's architecture retrained with an expectation-over-transformations channel: the same invertible network, the same payload sizes and training compute, but with the physical channel simulated by random affine warps, Gaussian noise, and JPEG instead of the trained transformation network. This is the strongest fair comparison because it isolates exactly my claimed ingredient — a learned differentiable channel model versus generic augmentation — while holding capacity, architecture, and budget fixed. Beating it means achieving reliable decoding (bit error rate below $10^{-3}$ after error correction) at 1 bit per pixel across the DCI grid of angles, distances, and lighting where the augmentation-trained variant either fails outright or only decodes at substantially reduced payload. Fixed Neural Network Steganography serves as the capacity reference (0.0% bit error up to 3 bpp, 35–39 dB container PSNR at 1 bpp) run per-image offline; I will measure its 10–20 second embedding latency and its physical-channel bit error rate to show both failure modes — too slow to deploy, too fragile to survive. An overt QR code printed on the same surface is the honest engineering alternative, run with equal error correction and surface area; it is beaten only on payload density under a perceptibility constraint, and I state plainly below what that does and does not prove. The key ablation within my own system is the band-placement constraint: the same trained network with payload routed into high-frequency subbands instead of low-frequency ones, which should reproduce the fragility collapse and confirm that survivability comes from band placement, not from capacity or architecture alone.
Why not just…?
Why not just print a QR code? For overt telemetry on a cooperative vehicle, a QR code plausibly works, and I will not claim otherwise. The proposal targets settings where the marking must be covert or appearance-preserving: provenance marks on printed AI-generated media in the spirit of InvisMark, or authentication that an adversary should not be able to locate and deface. A QR code fails the covert requirement by construction and its payload density per unit area at fixed perceptibility is orders of magnitude lower; the head-to-head experiment will quantify the density gap, and if overt marking is acceptable in a deployment, that deployment is simply not this paper's use case.
Why not just retrain RIIS with heavier augmentation? This is the real threat, and the strongest-competitor baseline above is designed to settle it rather than argue around it. The evidence favoring a learned channel model is DTA's "Renderer's Dilemma": naive differentiable proxies of physical rendering are inaccurate, which is why DTA trained a network to mimic a photorealistic non-differentiable simulator. But it is entirely possible that random augmentation over warps, noise, and compression covers enough of the physical distribution for a low-frequency payload. If the augmentation-trained baseline matches the transformation-network-trained system on the DCI grid and the physical print test, the contribution of this paper shifts honestly to the real-time distillation and the first physical-channel evaluation protocol, and I would report that result as such.
Why not just optimize per image offline, as FNNS does? Two measured failures rule it out for this regime. Embedding takes 10–20 seconds per image at 1–2 bpp and up to 3 minutes for the random-network variant, per Fixed Neural Network Steganography, which is incompatible with marking content at generation time or at frame rate; and FNNS itself reports that its payloads are destroyed by lossy operations, surviving JPEG only marginally at 0.1 bpp. The distillation experiment directly tests whether the feed-forward encoder matches iterative embedding quality at fixed payload — if it does not, that is a negative result I will report, since it would mean the real-time requirement and the high-capacity requirement genuinely conflict.
Why not just use radio for vehicle authentication? Dedicated short-range radio solves cooperative vehicle identity and I make no claim against it. The optical channel matters for the cases radio cannot reach: passive verification of uncooperative or foreign agents, printed media, and surfaces with no electronics. Nothing in the Lacuna file bears on this comparison, so I frame it as scope, not evidence.
Proposed Experiments
The first experiment is the channel sweep: train with the transformation network in the loop, then evaluate bit error rate across the full DCI grid — its 40 angles, 15 distances, and seven lighting/weather scenes, including the moving-observer condition where baseline detection accuracy falls to 29.66% — reporting payload density at which error-corrected decoding stays reliable, against the augmentation-trained and as-published RIIS baselines. The second is a physical validation at 1:10 scale following DTA's print-and-photograph protocol, measuring real sensor noise, lighting, and perspective effects against the simulated predictions; the transfer gap between simulation and physical capture is itself a reported result. The third is latency: embedder and extractor wall-clock on GPU and on an Eyeriss-class accelerator, with the pass condition set at 27 ms per frame, anchored to Themis's 36–59 fps envelope, and contrasted with IMPRESS purification at roughly 141 seconds per image — the number that disqualifies diffusion-based approaches from any real-time pipeline. The fourth is the adversarial scrubbing test, where the fragility direction becomes the threat model: after physical capture, an attacker applies IMPRESS-style diffusion purification and Feature-Distillation-style quantization (which restores adversarial accuracy from ~20% to ~90% on ImageNet) to strip the mark. Because the payload lives in the band purification preserves for semantic fidelity — the same reasoning by which IMPRESS restores style mimicry on GLAZE-protected images from ~42% back to 87% while leaving content intact — the mark should survive where high-frequency marks fail; an adaptive attacker using BPDA-style differentiation through the published transformation network is the strongest version of this test. The fifth is covertness: statistical steganalysis and a human conspicuousness study, contrasting with the loud geometric textures of DTA and RAUCA. The decisive check for the whole proposal is simple to state: at least 1 bit per pixel, decoded reliably, from a photograph, written and read in one frame — a combination no paper in this literature has ever achieved, and each existing line of work fails at exactly one of its four conditions.
References
- Feature Distillation: DNN-Oriented JPEG Compression Against Adversarial Examples
- IMPRESS: Evaluating the Resilience of Imperceptible Perturbations Against Unauthorized Data Usage in Diffusion-Based Generative AI
- DTA: Physical Camouflage Attacks using Differentiable Transformation Network
- Robust Invertible Image Steganography
- Fixed Neural Network Steganography: Train the images, not the network
- Semantic Deep Hiding for Robust Unlearnable Examples
- Toward Robust and Accurate Adversarial Camouflage Generation Against Vehicle Detectors
- DynamicPAE: Generating Scene-Aware Physical Adversarial Examples in Real-Time
- InvisMark: Invisible and Robust Watermarking for AI-Generated Image Provenance
- Benchmarking the Physical-world Adversarial Robustness of Vehicle Detection
- Real-Time Robust Video Object Detection System Against Physical-World Adversarial Attacks