A company that trains a self-driving perception system rarely curates every training image itself. Data arrives from outsourced training providers, third-party datasets, or cameras on its own vehicle fleet passively harvesting public roads — and a malicious party can plant a backdoor in this pipeline: a physical object or pattern (a patch of mud, a rust-like blemish on a sign) that, once seen a few hundred times in training data, makes the model misbehave in a chosen way whenever the pattern reappears at deployment. The uncomfortable fact, documented across the physical-attack literature, is that the most effective physical triggers are ones that look like natural damage: BadLANE's amorphous mud smears on lane markings (BadLANE) and AdvWT's corrosion and peeling-paint styles on traffic signs (Adversarial Wear and Tear). This creates a problem nobody has measured: the same model must learn natural wear in order to recognize weathered signs in the wild — adversarial training with wear styles raised damaged-sign accuracy from 90.1% to 95.4% in AdvWT — so a defense that deletes everything resembling the trigger also deletes the model's ability to cope with the real, damaged world.
This proposal studies that collision directly. Its core claim is that when a physical backdoor trigger is drawn from the distribution of natural damage, post-training purification faces a genuine entanglement problem — malicious shortcut channels and legitimate wear-robustness channels overlap — and that a style-counterfactual attribution can pull them apart: a channel is suspect only if its contribution to the attack is high while its contribution to recognizing genuinely worn signs is low. We will build the first testbed that measures backdoor removal and damaged-input accuracy on the same axis, test whether existing defenses collapse on it, and develop a defense that does not.
Core Claim
The closest existing work defines the gap from two sides. On the attack side, Clean-Label Physical Backdoor Attacks with Data Distillation implants physical triggers (sunglasses, tennis balls) into face and animal classifiers via dataset distillation, achieving nearly 100% attack success rate for its Gradient Matching variant while barely hurting clean accuracy — but it is an attack paper, evaluated only on pristine-image classification, with poor cross-architecture transfer and neutralization by Neural Attention Distillation. On the defense side, Adversarial Neuron Pruning Purifies Backdoored Deep Models (ANP) purifies an infected model with as little as 0.1–1% clean data, cutting attack success rate from nearly 100% to below 1% in 5 of 6 scenarios on CIFAR-10 at an average clean-accuracy cost of 1.47% — but every trigger it evaluates is digital and visually artificial, and the paper itself flags, without testing, that physical triggers sharing channels with natural objects would likely force severe clean-accuracy drops. Between them sits Adversarial Wear and Tear, which supplies the wear-style generator we reuse but studies only inference-time evasion, never backdoors or pipeline access. No cited work measures what happens when a purification defense meets a trigger that lives inside the natural-damage distribution — and no work reports the second accuracy axis (accuracy on naturally damaged clean inputs) that this question demands. That is the gap this proposal fills.
Our claim has three parts. First, empirically: we will test the unverified prediction that ANP-style pruning either fails on wear-camouflaged backdoors or destroys damaged-input accuracy — this is an open question, and if ANP turns out to work, the contribution shifts to the dual-axis benchmark and the stronger adaptive attack below. Second, mechanistic: we will quantify channel entanglement, measuring how much the "sleeper" channels that purification would prune overlap with the channels that carry wear robustness. Third, constructive: we propose style-counterfactual purification, a defense that scores each channel by separating its two roles before pruning.
Research Program
Testbed construction. Under the pipeline-access threat model — the defender receives a trained model from a pipeline it could not fully audit (outsourced training or passively harvested fleet data, as in ANP's MLaaS setting and as motivated by the closed-pipeline analysis in the backdoor-access literature) and holds only a small verified clean set — we implant backdoors whose triggers are natural-looking wear. For traffic-sign classification on GTSRB-HQ, we search adversarial style codes with the AdvWT generator, which reports attack success rates of 93.32% on GTSRB-HQ and 91.41% on its Hybrid dataset across ResNet, MobileNet, and ViT backbones, and human naturalness ratings of 3.66/5 versus 2.17 for laser-style attacks. For lane detection, we use BadLANE's mud-debris triggers on TuSimple and CULane with LaneATT and PolyLaneNet, where BadLANE reports an average attack-success-rate gain of +25.15% over BadNets/Blended baselines and sustained ASR above 90% under simulated rain and shadows. Poisoning budgets stay at or below 3% of training data, matching the 96%-success, sub-3%-poisoning regime of Physical Backdoor Attacks to Lane Detection Systems in Autonomous Driving. Evaluation applies REAP-style differentiable perspective and relighting transforms, since REAP shows synthetic evaluation overestimates physical effectiveness by 50–60 percentage points, plus a modest physical component (printed wear decals and mud patches under a distance/lighting grid, and optionally a Jetbot lane rig, where BadLANE manipulated trajectory in 95% of horizontal-view trials).
The entanglement measurement. For each infected model, we identify backdoor-carrying channels via ANP's own perturbation procedure, and wear-robustness channels via style-counterfactual stability: resample wear styles on clean damaged-sign images and keep channels whose activations stay stable while the class label is fixed. We then measure overlap. This tells us whether the trade-off is real and how sharp it is — the number nobody has reported.
Style-counterfactual purification. For each channel $i$, let $\Delta_{\text{trig}}(i)$ be the drop in the attacker's target-class logit on trigger-bearing images when channel $i$ is zeroed, and $\Delta_{\text{wear}}(i)$ the drop in the true-class logit on clean images whose wear style has been counterfactually resampled. We prune channels by descending ratio
$$ s_i = \frac{\Delta_{\text{trig}}(i)}{\Delta_{\text{wear}}(i) + \epsilon}, $$
with $\epsilon$ a small constant, so a channel is removed only when it powers the attack far more than it powers recognition of worn signs. This is the key difference from ANP: its min-max objective maximizes damage to the trigger response and is blind to what else a channel does; ours constrains the collateral term explicitly. Pruning is followed by a brief style-marginalized fine-tune on the small clean set, which restores any wear robustness lost. Finally, we include an adaptive attacker who optimizes the trigger to maximize overlap with legitimate wear features, and require the defense to degrade gracefully rather than collapse.
Baselines
The "do nothing" reference point is the infected model as delivered. Based on the attack papers above, we expect it to show attack success rates near 96–100% (Han et al. report 96% success with under 3% poisoning; CLPBA's Gradient Matching variant reaches nearly 100%) while retaining normal clean accuracy — PTB, for instance, maintains 96.33% clean accuracy alongside its attack. This baseline anchors the trade-off: a defense is only useful if it moves attack success rate down without dragging damaged-input accuracy with it.
The strongest existing competitor is ANP (Adversarial Neuron Pruning), which reports attack success rate reduced from nearly 100% to below 1% in 5 of 6 CIFAR-10 scenarios using 1% or even 0.1% clean data, at an average clean-accuracy cost of 1.47%. We will run ANP exactly as published, on the same infected models, with the same clean-data budgets (0.1%, 0.5%, 1%) and the same tuning effort, and score it on both axes. Beating it means achieving comparable attack suppression (attack success rate below roughly 5%) while either ANP fails outright on wear triggers or loses more than about 5 points of damaged-sign accuracy where we lose under 2. Two further defenses frame the frontier: plain fine-tuning on the clean set, which ANP's authors report causes 7%+ clean-accuracy drops when aggressive and which BadLANE found only marginally mitigated its amorphous triggers at a significant clean-accuracy cost; and AdvWT-style wear-augmented adversarial fine-tuning, which is the strongest known method on the second axis (it raised wild damaged-sign accuracy from 90.1% to 95.4%) and therefore the toughest test of whether robustness training alone removes backdoors.
The ablation isolating our key ingredient is the same pruning pipeline with the style-counterfactual denominator removed — that is, ANP-style trigger-response scoring alone — plus a variant with magnitude-only channel selection. If the ratio-based score does not outperform trigger-response-only selection specifically on damaged-sign accuracy at matched attack success rate, the entanglement story and the defense both fail, and we will say so.
Why not just…?
Why not just fine-tune or fine-prune on a small clean set? This is ANP, and it is the strongest version of the simple answer — but every cited measurement of it against physically natural triggers points the wrong way. BadDet+ found standard fine-tuning defenses largely insufficient across Faster R-CNN, FCOS, DINO, and YOLOv5 (BadDet+); CLPBA maintained high attack success rates even under full fine-tuning; and ANP itself concedes it was never evaluated on triggers sharing channels with natural objects. Our first experiment is precisely this baseline run on wear triggers, so the question is settled empirically rather than by argument — and if ANP wins, the proposal's framing changes rather than survives.
Why not just train or fine-tune with wear-style augmentation? This is the most serious alternative, because AdvWT's dual-use result shows wear-style adversarial training improves clean accuracy on damaged signs. But augmentation addresses the model's robustness to a style distribution, not a planted input-label correlation; nothing in the cited evidence suggests it removes an implanted shortcut, and BadLANE showed triggers mimicking natural noise survive augmentations that do not destroy lane content, while aggressive color-space augmentation degrades clean accuracy in its own right (Color Backdoor) and CLPBA bypassed MixUp outright. We will run this baseline at matched budget; our hypothesis is it recovers damaged-sign accuracy while leaving attack success rate high, possibly even hardening the trigger.
Why not just filter inputs at inference time? Certified patch defenses like PatchCleanser — which reports 83.9% clean and 62.1% certified robust accuracy on ImageNet against a 2%-pixel patch — need two masking rounds costing hundreds of milliseconds per image on high-end GPUs, which is unviable for 30-FPS vehicle perception. More fundamentally, they assume a small localized patch; Han et al. showed median filtering fails against physical cone triggers, and BadLANE's shapeless mud smears cannot be masked without masking legitimate lane structure. This alternative addresses evasion, not backdoors, and we treat it as out of scope.
Why not just retrain from scratch on verified data? For a fleet-scale pipeline the harvested data is the data; verified subsets are the 1% budget we already assume. Where full retraining is feasible, it is the ceiling any purification must approach, and we will report it as an oracle upper bound on a reduced-scale setting so the reader can see how much of the gap our method closes.
Proposed Experiments
The decisive sequence is as follows. First, implant wear-camouflaged backdoors in ResNet, MobileNet, and ViT sign classifiers on GTSRB-HQ, LaneATT and PolyLaneNet on TuSimple and CULane, and a Faster R-CNN detector on MTSD (where the FPN101 baseline reports 88.38 AP), and characterize attack success rate, pristine accuracy, and — the new axis — accuracy on the AdvWT wild damaged-sign set and BadLANE's degraded lane subsets, before and after REAP-style perspective and relighting transforms. Second, run ANP, fine-tuning, wear-augmented adversarial fine-tuning, and our method at matched clean-data budgets, reporting the full two-axis frontier: attack success rate versus damaged-input accuracy, with pristine accuracy as the third coordinate. Third, the entanglement analysis: overlap between ANP-flagged sleeper channels and style-stable wear channels, testing whether the observed trade-off tracks the overlap. Fourth, the ablations above, at matched attack success rate so collateral damage is compared fairly. Fifth, the adaptive attack: a trigger optimized to maximize channel overlap with legitimate wear, with the success criterion being graceful degradation (attack success rate rises, but only alongside a measurable rise in the attacker's constraint to look natural, which AdvWT's 3.66/5 naturalness protocol lets us verify with human raters). We will also check a training-set-cleansing defense, since Training set cleansing of backdoor poisoning by self-supervised representation learning reduces attack success rate from nearly 100% to below 5% on CIFAR-10 but pays a 2–3% clean-accuracy penalty from rejecting valid long-tail samples and was never tested on physical triggers — on a 400-class long-tail benchmark like MTSD, we expect that penalty to grow, which would itself be a finding. The single most important number in the whole program is the one no cited paper reports: attack success rate after defense, plotted against accuracy on naturally damaged clean inputs, for ANP and for ours, on the same infected model.
References
- Towards Robust Physical-world Backdoor Attacks on Lane Detection
- Adversarial Wear and Tear: Exploiting Natural Damage for Generating Physical-World Adversarial Examples
- Clean-Label Physical Backdoor Attacks with Data Distillation
- Adversarial Neuron Pruning Purifies Backdoored Deep Models
- Physical Backdoor Attacks to Lane Detection Systems in Autonomous Driving
- REAP: A Large-Scale Realistic Adversarial Patch Benchmark
- BadDet+: Robust Backdoor Attacks for Object Detection
- Color Backdoor: A Robust Poisoning Attack in Color Space
- PatchCleanser: Certifiably Robust Defense against Adversarial Patches for Any Image Classifier
- Training Set Cleansing of Backdoor Poisoning by Self-Supervised Representation Learning