Backdoors Ride on Innate Wiring: Predicting Physical Trigger Strength from Untrained Networks

A backdoor is a hidden behavior planted in a neural network during training. The model works normally until an input contains a specific "trigger," such as a reflection on glass or a sticker, and then it outputs whatever the attacker chose. This proposal tests one claim: much of what makes a trigger easy to plant and hard to remove is already fixed before training starts, by how strongly the units of a randomly initialized, untrained network happen to respond to it. Untrained convolutional networks are known to contain hundreds of units that respond selectively to faces and to dozens of object categories. Our question is whether poisoned training takes over these pre-existing detectors when it implants a backdoor, rather than writing new ones from scratch.

To test this properly, we need triggers that vary smoothly along physically meaningful dimensions and that we can optimize. A differentiable ray tracer provides that. It is a light-transport simulator whose output images can be differentiated with respect to scene settings such as surface roughness, light position, and glass thickness. With it, we can render glints, reflections, and interreflections that stay consistent as the camera moves. We can also pass gradients from an untrained network's units back into those physical settings. Doing so produces physically realizable triggers that are matched to the network's innate wiring. If the claim holds, it explains why natural reflection triggers survive pruning defenses. It also gives an architecture-level way to audit which physical phenomena are dangerous triggers without training a single poisoned model.

Core Claim

The core claim is that an untrained network's selectivity for a trigger predicts how readily that trigger can be backdoored. Specifically, it should predict the poison rate needed to reach high attack success and how much attack success survives pruning and fine-tuning. It should predict these better than low-level image statistics do, and comparably to or better than selectivity measured in a clean-trained model. The claim is also mechanistic: the neurons that carry the finished backdoor should overlap with the units that were trigger-selective at initialization, beyond what chance predicts.

We measure innate selectivity with a discriminability score. For unit $i$ of an untrained network, define

$$ S_i = \frac{\mu_i^{\text{trig}} - \mu_i^{\text{clean}}}{\sqrt{\tfrac{1}{2}\left(\sigma_i^{\text{trig}\,2} + \sigma_i^{\text{clean}\,2}\right)}}. $$

Here $\mu_i^{\text{trig}}$ and $\sigma_i^{\text{trig}}$ are the mean and standard deviation of the unit's activation over images carrying the trigger. $\mu_i^{\text{clean}}$ and $\sigma_i^{\text{clean}}$ are the same quantities over matched clean images. A trigger's innate score $A$ is the number of units per layer with significant $S_i$, averaged over random initialization seeds. This mirrors the face-selectivity index used in Face detection in untrained deep neural networks.

Grounding in Prior Work

Three findings motivate the claim. Face detection in untrained deep neural networks found about 250 significantly face-selective units in the last convolutional layer of every randomly initialized AlexNet, plus units selective for 39 ImageNet categories. It also found that training reduces the number of such units while sharpening the survivors, so learning acts partly by pruning an innate scaffold. A related finding of an innate number sense in untrained networks, Spontaneous generation of innate number sense in untrained deep neural networks, suggests this is a general property of random hierarchical wiring rather than something specific to faces.

Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks argues that backdoors live in "dormant" neurons that clean data barely activates, and removes them by pruning followed by fine-tuning. Neither line of work asks whether those dormant neurons are the innate detectors that clean training failed to prune. That is exactly our question.

Reflection Backdoor: A Natural Backdoor Attack on Deep Neural Networks (Refool) motivates the physical side. Reflection triggers reach over 90% attack success on GTSRB, BelgiumTSC, PubFig and ImageNet with under 3% clean-accuracy loss. They also retain 60–80% success after fine-tuning and pruning, where standard triggers fall below 20%. Refool builds its reflections in 2D, as a weighted blend of a background and a reflection image passed through blur or ghosting kernels. It offers no account of why reflections survive pruning. Our hypothesis supplies a candidate explanation: reflection statistics may align with innate, non-dormant units. We make this testable by varying the physics of the reflection.

Two physical-backdoor papers constrain the design. MOBA: A Material-Oriented Backdoor Attack Against LiDAR-Based 3D Object Detection Systems found that modeling material reflectance, through an angle-marginalized Oren–Nayar model, was necessary for its triggers to transfer to the physical world. It reports 93.5% attack success and a 41% margin over prior physical attacks. However, it works on active LiDAR returns and trained detectors. GaussTrap: Stealthy Poisoning Attacks on 3D Gaussian Splatting for Targeted Scene Confusion uses camera pose as the trigger in a radiance field, with no material or lighting model. Neither examines untrained networks, and neither optimizes triggers through camera light transport.

The rendering method is reused, not invented. We follow the Monte Carlo path tracing with in-loop denoising and detached importance sampling from Shape, Light, and Material Decomposition from Images using Monte Carlo Rendering and Denoising, which converges with 8–32 rays per pixel. We use the Disney principled material parameterization (albedo, roughness, metallic) from Inverse Path Tracing for Joint Material and Lighting Estimation. We render only small trigger scenes at classifier resolution, never full inverse reconstructions. This sidesteps the 3–5 hour per-object cost reported in Shape from Tracing.

Research Program

We first build a parameterized trigger family. A glossy laminated traffic-sign plate or a pane of glass in front of a face is rendered under an environment map, with an off-screen object whose reflection or specular glint lands on the surface. The controllable parameters are roughness, specular albedo, glass thickness (which produces ghosting), light direction, reflected-object position, and camera pose. A trigger is a setting of these parameters, rendered and composited onto GTSRB, PubFig, and CIFAR-10 images.

Two comparison families sit alongside the rendered triggers. The first is Refool's 2D blends. The second is a set of pixel-space patterns. Every trigger in all three families is matched on mean luminance change, RMS contrast, and power spectrum. This matching prevents "innate selectivity" from simply standing in for trigger salience.

Second, for each trigger we compute $A$ on untrained ResNet-18, VGG-16, and AlexNet across 20 random seeds. We then run poisoned training and record three outcomes. The first is the minimum poison rate that reaches 90% attack success. The second is attack success after fine-pruning. The third is the overlap between units carrying the backdoor and the units that were innately selective for the trigger. We identify backdoor-carrying units by the activation gap between triggered and clean inputs in the poisoned model. We measure overlap by intersection over union against a seed-matched permutation null.

Third, we run two causal tests. In the knockout test, we re-initialize the innately trigger-selective units at step zero, holding layer weight norms fixed, and then poison. The hypothesis predicts a higher required poison rate and lower survival after pruning. In the alignment test, we backpropagate from the untrained network's summed $S_i$ through the differentiable renderer into the scene parameters. This yields "innate-aligned" physical triggers, which we compare against random physical settings with matched contrast.

Finally, a secondary and conditional experiment asks whether the idea gives a defense. For models fine-tuned from a known public checkpoint, the defender knows the starting weights. Units whose response to non-task stimuli persisted or sharpened relative to that checkpoint, where clean training would have pruned it, are candidates for backdoor carriers. We prune these instead of the merely dormant units.

Baselines

The naive reference point is a clean-trained model with no poisoning, evaluated with triggers present. This sets the floor for "attack success" caused merely by distribution shift. Alongside it, we use random physical trigger settings that are rendered but not aligned. These give the expected strength of an arbitrary reflection.

The strongest existing attack in this natural-trigger family is Refool Reflection Backdoor. On GTSRB it reports over 90% attack success with under 3% clean-accuracy drop, and 60–80% success after fine-tuning and pruning. We rerun it on the same architectures, poison budgets, training schedule, and contrast-matched trigger pools as our rendered triggers. Our aligned triggers beat it only if they reach 90% success at a lower poison rate, or retain clearly more success after fine-pruning, at equal clean accuracy and equal perceptual contrast.

On the defense side, we use two baselines. The first is Fine-Pruning Fine-Pruning, which reduced attack success from 100% to 0% on YouTube Faces but only from about 99% to about 30% on U.S. traffic signs. The second is STRIP STRIP, which reports 0% false acceptance and false rejection in most CIFAR-10 and GTSRB cases. Both get the same clean calibration set and the same tuning budget as our persistence-based pruning. Our defense counts as better only if it lowers residual attack success below Fine-Pruning's level at equal clean-accuracy loss.

For the predictive claim, the competing predictors are low-level trigger statistics and the selectivity of the same units in a clean-trained model. The innate score must beat the first and at least match the second in rank correlation with the required poison rate. Two ablations isolate the key ingredient. The knockout at initialization tests the mechanism. Rendered triggers with the alignment objective removed test whether alignment, rather than rendering itself, drives any gain.

Why not just…?

Why not just use Refool's 2D blends instead of a ray tracer? For the core predictive and knockout claims, 2D triggers may suffice. That is why they form a full comparison arm rather than a strawman. If innate scores predict backdoor strength for 2D blends just as well, the claim survives and the renderer's role shrinks to the physical-realizability experiments. The renderer remains necessary for two things 2D blending cannot do. First, it produces triggers whose appearance changes consistently with camera pose and lighting; MOBA's finding is that ignoring reflectance physics broke physical transfer. Second, it lets gradients flow into physical settings, which are the only variables an attacker in the real world controls. We test this by evaluating attack success on held-out rendered camera poses and light directions. On those, 2D blends should degrade and rendered triggers should not.

Why not just optimize a pixel pattern to maximize innate selectivity? In purely digital attack success, a pixel pattern will probably win, and we expect to report that honestly. The question here is different: whether naturally occurring physical phenomena align with innate scaffolds, and whether that alignment explains their survival under pruning. We run pixel-optimized triggers alongside physical ones. The decisive comparison is survival under viewpoint and lighting change and under fine-pruning, not raw success on fixed images.

Why not just say trigger salience explains everything? This is the most serious threat. Highly contrasted, spectrally distinctive triggers might both excite random units and be easy to learn, with no special role for innate wiring. We address it in three ways. Triggers are matched on contrast and spectrum. Low-level statistics are entered as a competing predictor. Most directly, the knockout experiment holds the trigger fixed and changes only the initial units. If knockout leaves the required poison rate unchanged, the mechanistic claim fails, even if the correlation holds.

Why not just measure selectivity in the clean-trained model? If clean-trained selectivity predicts as well, our contribution reduces to "already-selective units get reused," which is less interesting. Two predictions separate the stories. Training prunes innate units, as Face detection in untrained deep neural networks found. So triggers selective at initialization but pruned by clean training should still be easy to poison, because poisoning would rescue those units. The innate score, unlike a clean-trained score, can also be computed before any training.

Proposed Experiments and Decisive Checks

The first decisive check is a rank correlation between the innate score $A$ and the required poison rate. It is computed across roughly 200 triggers spanning the rendered, 2D, and pixel families, on GTSRB with ResNet-18, and replicated on PubFig and CIFAR-10. The result is reported with the low-level and clean-trained predictors entered jointly in a regression.

The second is the unit-overlap test against a permutation null. The third is the knockout at initialization. A rise in the required poison rate together with a drop in survival after fine-pruning would support the mechanism. No change would falsify it.

The fourth is the alignment test. Renderer-optimized triggers are compared with random physical settings and with Refool on GTSRB. The comparison uses required poison rate, attack success on held-out rendered poses and lights, and success after Fine-Pruning, all at matched clean accuracy. The Refool comparison targets its reported 60–80% post-defense success.

Finally, the persistence-based pruning defense is compared with Fine-Pruning and STRIP on the same poisoned models. There is no reported result on Lacuna for any defense against rendered, view-consistent physical triggers, so these numbers will be new. Results will be reported per architecture, because the innate scaffold, and therefore the effect, may depend strongly on depth and receptive-field size.