Every camera sensor is a photon integrator: the intensity a pixel reports is roughly ambient light plus whatever signal the scene carries, so the usable contrast of any engineered pattern shrinks as $\text{Contrast} \propto I_{\text{adv}} / I_{\text{ambient}}$. The machine-learning literature treats this as a passive, one-directional fact: bright daylight washes out projector and LED attacks, and dim conditions wash out the defenses that try to spot them. Both sides treat ambient light as a scalar noise floor to be endured. Biology takes the opposite view: visual cortex divides each neuron's response by a pooled estimate of overall activity — divisive normalization — and theory shows this actively stabilizes recurrent circuits precisely when external input grows stronger, extending the stable operating range by more than 100% over linear-gain networks (Stabilization of recurrent neural networks through divisive normalization). Nobody has asked what happens if a camera vision system does the same thing.
This proposal ports ambient-driven divisive normalization into a recurrent video perception network, so that rising illumination engages gain control that expands the network's dynamical stability margin instead of eroding its statistics. The same nonlinear dynamics then yield a second, unexploited capability: as the network's latent state approaches a bifurcation — for example when a high-frequency LED flicker attack begins driving the sensor — it exhibits critical slowing down, a measurable rise in the autocorrelation and variance of the latent state that the stabilization theory quantifies as $\Delta_{csd} \sim 0.09$ before instability (Stabilization of recurrent neural networks through divisive normalization). That signature can serve as an early-warning attack detector built from the model's own dynamics, requiring no separate detection network.
Core Claim
Ambient illumination and optical attacks should be modeled as measured inputs driving a nonlinear dynamical system, not as additive noise on a static feedforward pipeline. Three observations motivate this. First, the stabilization theory proves that divisive normalization keeps recurrent networks stable even when the recurrent weight spectral radius exceeds one, and that stronger external inputs actively expand the stability boundary (Stabilization of recurrent neural networks through divisive normalization) — exactly the property a perception system needs as lux levels swing by three orders of magnitude. Second, physical optical attacks already exploit temporal dynamics: rolling-shutter attacks modulate RGB LEDs above 60 Hz so the camera's row-by-row exposure converts temporal waveforms into spatial adversarial stripes, achieving an 84% targeted attack success rate on ImageNet with ResNet-101 at confidence often above 70% (Invisible Perturbations: Physical Adversarial Examples Exploiting the Rolling Shutter Effect); fluorescent patches achieve 98.31% success across ten traffic-sign models at 200 lux (The Fluorescent Veil: A Stealthy and Effective Physical Adversarial Patch Against Traffic Sign Recognition). Yet every one of these attacks treats the victim as a static network, and every defense treats illumination as a per-frame nuisance. Third, the input-driven dynamical modeling literature has already solved the analogous "conflation" problem in another domain: BRAID shows that explicitly modeling external drivers prevents mistaking input-driven shifts for internal dynamical transitions, reaching a behavioral forecasting correlation of 0.8042 versus 0.7453 for linear models on primate motor cortex (BRAID: Input-driven Nonlinear Dynamical Modeling of Neural-Behavioral Data). We apply the same move to camera vision: ambient illumination and the camera's own auto-exposure state are external drivers, and conflating them with scene content is a dynamical-modeling failure, not just a signal-to-noise problem.
Research Program
The core artifact is a recurrent video-classification backbone (a recurrent or continuous-depth network processing short frame bursts) in which the recurrent feedback passes through a divisive normalization stage whose pool includes an explicit ambient-intensity estimate:
$$r_i(t) = \frac{h_i(t)}{\sigma + \alpha\,\hat{I}_{amb}(t) + \beta \sum_j |h_j(t)|}$$
Here $h_i(t)$ is the activity of recurrent unit $i$, $\hat{I}_{amb}(t)$ is a running ambient-intensity estimate read from the camera's exposure and gain metadata (a cheap, physically grounded proxy available on any real sensor), $\sum_j |h_j(t)|$ is the pooled recurrent activity, and $\sigma$, $\alpha$, $\beta$ are learned scalars. As ambient light rises, the denominator grows and recurrent gain falls — the mechanism by which ORGaNICs-style normalization expands the stability boundary under stronger input. This directly addresses the known limitation of the stabilization theory, which is proven only for symmetric Gaussian Orthogonal Ensemble matrices (Stabilization of recurrent neural networks through divisive normalization): we test whether it survives learned, asymmetric weights in a real perception pipeline.
Two further components. First, we add attractor regularization in the spirit of RMAN, which forces representations toward discrete point attractors with an auxiliary energy loss and improved CIFAR-10 error from 7.95% to 7.61% for ResNet-20 (Reinforcing Neural Network Stability with Attractor Dynamics); our hypothesis is that its benefit is far larger under optical perturbation than in the clean setting where it was originally measured. Second, we build the early-warning detector: a lightweight online statistic tracking the autocorrelation and variance of the latent state near its fixed point, firing when critical slowing down precedes a bifurcation. The geometry here follows the attractor-inversion result that adversaries manipulate dynamical systems by moving fixed points rather than individual outputs — shifting a fixed point from $-0.24$ to $+1.11$ in human decision models (Attractor Inversion: A Geometric Account of Adversarial Manipulation in Human Decision-Making) — and we test whether optical flicker produces the analogous signature in a visual network. One honest scope limit: when raw sensor saturation physically clips the signal (fixed exposure under direct 100,000-lux sunlight), no architecture helps; our target regime is the enormous middle range where auto-exposure adapts and the problem is that exposure and gain changes destabilize the model's internal statistics and the timing of shutter-dependent attacks.
Proposed Experiments
The central experiment is an attack-success-rate-versus-illuminance sweep across three attack channels taken from the literature. In simulation, we place GTSRB and CTSRD traffic signs under digitally controlled illumination from 200 lux upward with fluorescent-patch adversaries in the FIPatch setting, whose reported baseline is a 98.31% attack success rate at 200 lux across ten models including YOLOv3 and Faster R-CNN (The Fluorescent Veil). We replicate the rolling-shutter channel on ImageNet, where the reported baseline is an 84% targeted success rate on ResNet-101 with the attack strictly dependent on exposures faster than 1/750 s (Invisible Perturbations). The projector channel follows the device-aware attack, whose authors report a 14% average similarity drop from simulation to physical projection and mitigation to nearly 100% success via grayscale multi-view optimization (Device-aware Optical Adversarial Attack for a Portable Projector-camera System); we adopt their grayscale physical modeling to keep our sim-to-real gap small. The physical rig is a camera with a rolling shutter, an LED modulator above 60 Hz, and programmable lighting reproducing the Invisible Perturbations hardware setup, swept from 200 lux to bright daylight. For each attack channel and each lux level we report attack success rate, clean accuracy, and the area under the success-rate-versus-lux curve — the primary metric, since single-lux numbers hide where each system fails. The file reports no number for rolling-shutter attack success under bright sunlight on Lacuna, so the sweep itself will establish that curve. For the early-warning detector we report detection AUC and lead time in frames before misclassification; no reported result found on Lacuna for dynamical early-warning attack detection, which is precisely the open territory. Sanity checks anchor the pipeline in known numbers: CIFAR-10 and CIFAR-100 clean error against the RMAN references (7.95% ResNet-20, 25.46% LM-ResNet-110), and per-frame illumination normalization on Ambient6K against the PromptNorm reference of 22.116 dB PSNR (PromptNorm: Image Geometry Guides Ambient Light Normalization). We also test the mechanism directly by estimating the effective spectral radius of the recurrent map as ambient input rises, expecting the normalized network to hold it below one where a fixed-gain counterpart exceeds it, mirroring the theory's reported 100% stability extension.
Baselines
The do-nothing reference point is the standard feedforward victim at auto-exposure: ResNet-101 for the ImageNet rolling-shutter channel and the YOLOv3/Faster R-CNN family for traffic signs, the exact victims used in Invisible Perturbations and The Fluorescent Veil. Their reported numbers — 84% targeted success and 98.31% at 200 lux respectively — define the failure floor. The strongest existing defense is NutNet, a 5,000-parameter reconstruction autoencoder that masks adversarial patches as out-of-distribution artifacts, improving defense performance 2.4-fold for hiding attacks and 4.7-fold for appearing attacks at 8% latency overhead and 0.4% clean-accuracy cost (I Don't Know You, But I Can Catch You: Real-Time Defense against Diverse Adversarial Patches for Object Detectors). It is the strongest competitor because it is the only cited defense with quantified gains across many detectors (INRIA, COCO2014, KITTI, APRICOT), and — critically — its own authors acknowledge it fails in low light where patch contrast collapses, which is the regime our mechanism is built for. We run it on identical data, identical compute, and an identical tuning budget (matched hyperparameter grid search), and we count ourselves as beating it only if we reduce the area under the attack-success-versus-lux curve by at least half across the full 200-lux-to-daylight range while staying within its overhead envelope (no more than 8% latency, no more than 1% clean-accuracy loss). The upstream-normalization baseline is PromptNorm applied before a standard network, using its Ambient6K reference of 22.116 dB; we give it the same tuning budget, and beating it means matching or exceeding its clean illumination robustness while additionally resisting temporal attacks it cannot see. The key ablation replaces $\hat{I}_{amb}(t)$ in the normalization pool with a constant, keeping the architecture identical: if the ambient-driven division is truly the active ingredient, this variant should collapse back toward fixed-gain behavior at the illumination extremes. A second ablation removes the attractor energy loss, isolating its contribution from the normalization's.
Why not just…?
Why not just train with brightness and contrast augmentation? This is the most serious alternative, and we run it as a trained baseline rather than arguing it away: Expectation Over Transformation with lux-swept illumination jitter, given the same tuning budget. The existing evidence is not encouraging for it — every standard model in the cited attacks was trained with conventional pipelines and still fell: 84% targeted success on ResNet-101 and 98.31% across ten traffic-sign models (Invisible Perturbations, The Fluorescent Veil), and the 14% sim-to-real similarity drop reported for projector attacks shows digital modeling of photonic hardware is itself lossy (Device-aware Optical Adversarial Attack). Still, augmentation could plausibly close much of the clean-accuracy gap at high lux. If it does, our claim narrows honestly to the two things augmentation cannot provide: performance in the low-light regime where contrast-based cues vanish, and the dynamical early-warning signal, which is a property of the state space rather than the training distribution.
Why not just fix illumination upstream with a relighting model? PromptNorm already achieves 22.116 dB on Ambient6K using scene geometry to normalize lighting (PromptNorm). But it is an offline, frame-by-frame restoration model with no temporal memory, so it cannot act on modulation above 60 Hz — the rolling-shutter channel lives entirely above the flicker-fusion threshold — and it supplies no gain stabilization inside the classifier. The head-to-head experiment on the same sweep settles this: if pre-processing plus a standard network matches our backbone everywhere, the simpler pipeline should win and we will say so.
Why not just use a Kalman filter to track and subtract ambient illumination? A linear tracker is convex, cheap, and tempting. But linear systems cannot represent the phenomena at stake: they possess a single attractor and cannot hold multiple stable states or pass through bifurcations, which is why a locally linear dynamical model achieved a prediction error of 0.244 on the Wong-Wang decision task where a nonlinear velocity-field model achieved 0.002 (Interpretable Nonlinear Dynamic Modeling of Neural Trajectories), and why linear forecasting reached a correlation of 0.7453 against 0.8042 for the input-driven nonlinear BRAID (BRAID). The exposure controller's coupling of gain, shutter, and illumination is itself nonlinear feedback, and the flicker attacks live precisely in that regime. We will run the Kalman variant anyway as a baseline; its expected failure mode is a lagging estimate that misreads exposure-driven shifts as scene changes — the conflation problem BRAID identified.
Why not just use NutNet and be done? As the strongest existing defense we treat it as the primary competitor, not a dismissal. Its acknowledged low-light failure and its lack of any temporal memory make it structurally unable to handle the dark end of the sweep or high-frequency modulation, but at bright lux it may well win. The decisive comparison is the full lux-swept area metric under matched compute; if NutNet dominates at high lux while we dominate below roughly 1,000 lux and on early warning, the combined result is a regime map of which defense works where — itself a contribution the field currently lacks.
Decisive Checks
Three results would make or break the claim. First, the clean-accuracy-versus-lux curve: the normalized recurrent backbone should hold accuracy within one percent across the sweep where the fixed-gain ablation and NutNet degrade at the extremes. Second, the mechanism check: the estimated effective spectral radius of the recurrent map should stay below one as ambient input rises in the normalized network and cross it in the ablation, matching the theory's stabilization prediction on asymmetric learned weights for the first time. Third, the early-warning check: on flicker onset, the critical-slowing-down statistic should rise measurably before misclassification, with a detection AUC above 0.9 and a lead time of at least three frames; if the latent state shows no such signature, the detection claim fails cleanly even if the robustness claim survives. Together these tests decide whether ambient light is a passive noise floor or an input that a properly normalized dynamical system can turn to its own advantage.
References
- Stabilization of recurrent neural networks through divisive normalization
- Invisible Perturbations: Physical Adversarial Examples Exploiting the Rolling Shutter Effect
- The Fluorescent Veil: A Stealthy and Effective Physical Adversarial Patch Against Traffic Sign Recognition
- BRAID: Input-driven Nonlinear Dynamical Modeling of Neural-Behavioral Data
- Reinforcing Neural Network Stability with Attractor Dynamics
- Attractor Inversion: A Geometric Account of Adversarial Manipulation in Human Decision-Making
- Device-aware Optical Adversarial Attack for a Portable Projector-camera System
- PromptNorm: Image Geometry Guides Ambient Light Normalization
- I Don't Know You, But I Can Catch You: Real-Time Defense against Diverse Adversarial Patches for Object Detectors
- Interpretable Nonlinear Dynamic Modeling of Neural Trajectories