Every physical adversarial attack deployed in the real world — a patch printed on a road sign, a marking painted on asphalt, a texture on a T-shirt — is made of matter, and matter weathers. Yet the adversarial-stealth literature treats physical patches as if they were timeless: a pattern is optimized in a generative model's latent space, printed once, and evaluated the same day. Meanwhile, a completely separate materials literature has spent years learning how to stabilize surfaces against exactly the ambient stressors — moisture, UV, oxidation — that will erode those printed patches. Neither literature has ever been pointed at the other. This proposal argues that the contest between them is measurable, and builds the first instrumented evaluation of it: an accelerated-weathering test bed for naturalistic adversarial patches on sign-grade substrates, with and without hydrophobic stabilizing coatings, paired with a differentiable weathering model that can be inserted into the attack's latent optimization.
The immediate motivation is the strongest existing stealth attack, Adversarial Wear and Tear (AdvWT), which disguises adversarial perturbations as natural rust, peeling paint, and fading on outdoor signs and achieves a 93.32% attack success rate on GTSRB-HQ across ResNet, MobileNet, and ViT victims, with a human naturalness rating of 3.66/5 Adversarial Wear and Tear: Exploiting Natural Damage for Generating Physical-World Adversarial Examples. AdvWT is purely digital: its "damage style code" is a learned image-texture manifold with no model of the chemistry that actually produces rust or the coatings that prevent it. The open questions it leaves are physical. Does a printed damage-disguised patch survive months of outdoor weathering, or does the very weathering it imitates destroy its adversarial spectral signature? And does the materials-science playbook of intrinsic stabilization — hydrophobic topcoats, defect passivation — protect the attack (making the patch persistent) or expose it (making synthetic damage conspicuous on a substrate that never naturally degrades)? No reported result found on Lacuna for attack success of any physical adversarial method under real or accelerated weathering.
Core Claim
The claim is that physical persistence and perceptual stealth of environmental-disguise attacks are jointly governed by substrate chemistry, and that this coupling can be quantified with existing tools from both fields. Concretely, we hypothesize three measurable regimes. On an unstabilized substrate, a damage-disguised patch decays in effectiveness under weathering, and — worse for the attacker — its decay is itself natural, so the attack quietly dies while remaining plausible. On a hydrophobically stabilized substrate (contact angle above roughly 90°, in the regime demonstrated by PMMA passivation at 93.8° and dopant-free chlorinated films at 94° Red Low-threshold Amplified Spontaneous Emission of Perovskite Films Achieved by A-site Substitution and Defect Passivation, Chlorine-Substituent Regulation in Dopant-Free Small-Molecule Hole-Transport Materials Improves the Efficiency and Stability of Inverted Perovskite Solar Cells), moisture- and oxidation-driven pattern erosion slows sharply, but the physically plausible space of damage textures collapses: moisture-blistered or corroded patterns become physically inconsistent with the surface's known chemistry, and a consistency checker can flag them. The interesting middle regime — and the one we most want to characterize — is a substrate whose stabilization is partial or patchy, where the attack must optimize within the physically realizable damage manifold, trading stealth margin for persistence.
Research Program
Three components, each reusing an existing technique from one of the two directions.
First, a physical weathering test bed. We fabricate AdvWT-style adversarial patches (StarGAN-v2 damage style codes, Monte Carlo latent sampling, following the published recipe) on aluminum traffic-sign blanks with standard retroreflective sheeting, in three substrate variants: bare, PMMA-topcoated, and coated with a chlorinated hydrophobic molecular layer, the last two chosen as direct analogues of the stabilization chemistries reported in Red Low-threshold Amplified Spontaneous Emission of Perovskite Films Achieved by A-site Substitution and Defect Passivation (PMMA carbonyl passivation raising ambient stability from 60% to 93% signal retention under 40% humidity) and Chlorine-Substituent Regulation in Dopant-Free Small-Molecule Hole-Transport Materials Improves the Efficiency and Stability of Inverted Perovskite Solar Cells (30-day ambient retention of 93.8% at contact angle 94°). We expose them to accelerated weathering — cycled UV, 40–95% relative humidity, and water spray, the standard cyclic-stress protocol that the stabilization literature itself flags as the unsolved case, since most passivations are validated only under static conditions. At fixed intervals we re-capture the panels under controlled illumination and re-measure attack success rate against the same victim classifiers, producing ASR decay curves per substrate.
Second, a differentiable weathering operator. We fit a compact model $W_t(\cdot\,;\phi)$ that maps a printed pattern's appearance to its appearance after weathering duration $t$, parameterized by substrate descriptors $\phi$ (measured contact angle, coating thickness, pigment type), trained on the captured time series from the chamber. This is the piece missing from all latent-space attack pipelines: Adversarial Wear and Tear and CLIP2Protect both optimize purely statistical latents with no physical constraint, and Naturalistic Physical Adversarial Patch for Object Detectors handles only the camera-side transform (Expectation of Transformation over viewpoint and lighting), not substrate-side decay. Validation of $W$ is a prediction task: forecast held-out weathered captures and report image-reconstruction error.
Third, a weathering-aware attack and a physics-consistency defense, which are the two sides of the core claim. The attack optimizes the damage style code $s$ so the pattern stays effective and naturalistic after simulated weathering:
$$\max_{s}\; \mathbb{E}_{t \sim [0,T],\,\phi}\Big[\mathrm{ASR}\big(f\,(W_t(G(z,s);\phi))\big)\Big] \;-\; \lambda\, d_{\text{phys}}(s, \phi)$$
where $G(z,s)$ is the generator rendering damage texture $s$ onto sign image $z$, $f$ is the victim classifier, $W_t$ is the fitted weathering operator, and $d_{\text{phys}}$ penalizes damage patterns inconsistent with substrate chemistry (for example, moisture-stained textures on a surface whose measured contact angle of ~94° makes water ingress implausible, in the spirit of the contact-angle-stability couplings documented across; Red Low-threshold Amplified Spontaneous Emission of Perovskite Films Achieved by <italic>A</italic>-site Substitution and Defect Passivation; Red Low-threshold Amplified Spontaneous Emission of Perovskite Films Achieved by <italic>A</italic>-site Substitution and Defect Passivation and Chlorine-Substituent Regulation in Dopant-Free Small-Molecule Hole-Transport Materials Improves the Efficiency and Stability of Inverted Perovskite Solar Cells Chlorine-Substituent Regulation in Dopant-Free Small-Molecule Hole-Transport Materials Improves the Efficiency and Stability of Inverted Perovskite Solar Cells). The defense is a binary consistency detector trained on (damage texture, substrate descriptor) pairs, asked to flag physically impossible weathering — the converse of the attack. This closes the loop that AdvWT opened but never tested: its authors showed that adversarial training on synthetic wear and tear lifts accuracy on real damaged signs from 90.1% to 95.4% Adversarial Wear and Tear: Exploiting Natural Damage for Generating Physical-World Adversarial Examples; we ask whether substrate-level stabilization, which removes the natural damage distribution itself, is a complementary and orthogonal defense.
Baselines
The "do nothing" reference point is the bare, unstabilized AdvWT patch: printed on standard sign sheeting, no coating, evaluated with the published digital pipeline on GTSRB-HQ and the 4,529-image Wear-and-Tear dataset against ResNet, MobileNet, and ViT. This establishes the unweathered attack strength and the natural decay rate that any coating or any weathering-aware optimization must beat. The strongest existing competitor is AdvWT itself Adversarial Wear and Tear: Exploiting Natural Damage for Generating Physical-World Adversarial Examples, which reports 93.32% ASR on GTSRB-HQ, 91.41% ASR on its Hybrid dataset, 3.66/5 human naturalness, and the 90.1%→95.4% adversarial-training gain. We will re-implement it faithfully (same generator, same Monte Carlo sampling budget, same victim models, same compute for latent search) and run it through the identical weathering protocol, so the comparison isolates exactly our additions — the weathering operator inside the objective and the stabilized substrates outside it. AdvWT's published numbers are pre-weathering and digital only; no reported result found on Lacuna for its ASR after physical weathering, so our first chamber measurement of the re-implementation is itself a contribution. The key ablation is our weathering-aware attack run on the unstabilized substrate and the vanilla AdvWT run on the stabilized substrate: this 2×2 disentangles whether any observed persistence gain comes from the coating, from the weathering-aware optimization, or requires both. Beating the baseline means, for the attack side, a strictly higher area under the ASR(t) curve over the weathering horizon at equal or better human naturalness (non-degraded naturalness score relative to AdvWT's 3.66/5), and for the defense side, a consistency-detector AUROC on stabilized substrates that is materially higher than on unstabilized ones, at a fixed false-positive rate on genuinely damaged unstabilized signs.
Why not just…?
Why not just add digital weather augmentation during classifier training? This is the simplest defense, and it partially works: Adversarial Wear and Tear itself showed adversarial training on its own worst-case synthetic damage lifted real-damaged-sign accuracy from 90.1% to 95.4%, which random corruption filters do not achieve. But augmentation only covers the damage distribution the generator can synthesize; it cannot cover the physically realizable manifold on a substrate it knows nothing about, and Beware of Road Markings shows models stay vulnerable to structural contextual traps (a Mean Relative Shift Ratio of 1.507 on KITTI across eight depth models) even when standard lighting and transformation augmentation is accounted for. Our experiment settles it directly: we will run the augmented-training defense against both vanilla and weathering-aware attacks on stabilized and bare substrates, and report whether augmentation alone closes the gap our physics constraints expose.
Why not just laminate the patch (attacker) or the sign (defender) with standard UV acrylic? Off-the-shelf lamination is the obvious cheap answer for durability on both sides, but it fails for opposite reasons. For the attacker, a glossy laminate introduces specular reflections and glare that the camera-angle robustness optimization (Expectation of Transformation) was never designed to absorb, breaking the very capture robustness that physical attacks depend on, a limitation noted for road-marking attacks in Beware of Road Markings; a matte intrinsic hydrophobic coating like the PMMA and chlorinated layers cited above adds ~90°+ contact-angle protection without an optically distinct surface layer. For the defender, external lamination addresses nothing at the interface — the materials literature's standing complaint, evidenced by grain-boundary oxidation and uncoordinated-ion defects proceeding under external seals in Regulation of Low-dose Stannous Iso-octanoate for Two-step Prepared Sn-Pb Alloyed Perovskite Solar Cells. We will include a laminated arm in the chamber protocol so this is measured rather than asserted.
Why not just constrain perturbations with a small Lp-norm bound instead of all this material machinery? Norm-bounded digital noise is the classical stealth recipe, and it collapses in the physical settings we care about: printed norm-bounded patches read as high-frequency noise to humans, earning an 8.3% naturalness rating against 75% for latent-optimized naturalistic patches in the human study of Naturalistic Physical Adversarial Patch for Object Detectors, and fine pixel-level perturbations rarely survive the print–camera optical transfer at all. The environmental-disguise regime is the one where stealth and physical persistence are both achievable, which is exactly why it needs a materials-level treatment.
Proposed Experiments and Decisive Checks
Experiment one measures ASR decay. Panels carrying identical AdvWT patches are weathered on the three substrate variants; attack success against ResNet, MobileNet, and ViT is re-measured at each capture point on GTSRB-HQ-classified content, giving ASR half-life per substrate. The stabilization literature's own numbers set the expectation for the coating effect: PMMA-style topcoats retained 93% of optical signal after 120 minutes of pulsed irradiation at 40% humidity versus 60% for unpassivated films Red Low-threshold Amplified Spontaneous Emission, and hydrophobic molecular layers held 93.8% efficiency after 30 days of ambient aging Chlorine-Substituent Regulation. If ASR half-life on coated substrates scales similarly relative to bare ones, the persistence half of the core claim holds. Experiment two measures stealth decay in the same protocol: human raters score naturalness on the AdvWT 1–5 protocol (its 3.66/5 versus AdvLaser 2.17 and AdvShadow 1.66 is the anchor Adversarial Wear and Tear: Exploiting Natural Damage for Generating Physical-World Adversarial Examples), and the physics-consistency detector scores every capture. The decisive outcome for the defense half of the claim is a large AUROC gap between stabilized and bare substrates, because on a stabilized surface any damage pattern must be synthetic or mechanically inflicted, both of which are detectable events. Experiment three closes the loop with the weathering-aware attack of the equation above: it succeeds only if it beats vanilla AdvWT's area under the ASR curve at matched naturalness on at least one substrate, and its failure modes are informative either way — if the physically plausible damage manifold on a 94°-contact-angle surface is too narrow to contain an effective attack, that is itself the strongest defense result in the proposal, and the first quantitative evidence that infrastructure stabilization closes a physical adversarial attack surface.
Two limitations bound the claims. The weathering operator is empirical, fitted to our chamber and substrates, not a first-principles kinetics model; we validate it only by held-out forecast accuracy, and we state its transfer range explicitly. And all physical attacks are confined to our own laboratory panels with no deployment on public infrastructure; the attacker-side results are reported as robustness findings for defenders, in the same dual-use framing AdvWT itself used when it converted its attack into a 90.1%→95.4% training gain.
References
- Adversarial Wear and Tear: Exploiting Natural Damage for Generating Physical-World Adversarial Examples
- Red Low-threshold Amplified Spontaneous Emission of Perovskite Films Achieved by <italic>A</italic>-site Substitution and Defect Passivation
- Chlorine-Substituent Regulation in Dopant-Free Small-Molecule Hole-Transport Materials Improves the Efficiency and Stability of Inverted Perovskite Solar Cells
- CLIP2Protect: Protecting Facial Privacy Using Text-Guided Makeup via Adversarial Latent Search
- Naturalistic Physical Adversarial Patch for Object Detectors
- Red Low-threshold Amplified Spontaneous Emission of Perovskite Films Achieved by <italic>A</italic>-site Substitution and Defect Passivation
- Chlorine-Substituent Regulation in Dopant-Free Small-Molecule Hole-Transport Materials Improves the Efficiency and Stability of Inverted Perovskite Solar Cells
- Adversarial Wear and Tear: Exploiting Natural Damage for Generating Physical-World Adversarial Examples
- Beware of Road Markings: A New Adversarial Patch Attack to Monocular Depth Estimation
- Regulation of Low-dose Stannous Iso-octanoate for Two-step Prepared Sn-Pb Alloyed Perovskite Solar Cells