Backdoors Do Not Transfer Because They Are Sharp: A Curvature Account and a Transfer-Deficit Audit

Two well-documented facts about neural networks point in opposite directions. Adversarial perturbations, meaning tiny input changes that flip a model's prediction, often fool other models trained on the same task. Backdoor triggers do not. A backdoor trigger is a hidden pattern planted during training that forces an attacker-chosen output, and such triggers fail to move the features of an independently trained model. We propose that one geometric fact explains both. The decision boundary is the surface in input space where the model's prediction changes. A backdoor carves a thin, sharply curved pocket into this surface that exists only in the poisoned model, while ordinary adversarial directions follow broad, nearly flat regions of the boundary that many models share. If this is right, a perturbation that reaches the backdoor's target class by riding the shortcut will be sharp and will not transfer, whereas perturbations toward clean classes transfer at their usual rate.

This gives a practical audit. Take a suspect model you can only query for labels. Search for small perturbations that push inputs into each class. Then measure how much worse those perturbations transfer to a few clean reference models than perturbations crafted on the references themselves. A class with an abnormal "transfer deficit" is a likely backdoor target. The test compares each class against clean references rather than against the other classes. That matters because the leading hard-label detectors assume that only a minority of classes are infected, and they weaken exactly when many classes are.

Core Claim

The claim has two parts. The first is mechanistic. Among targeted adversarial perturbations found on a backdoored model, those toward the attacker's target class land at boundary points with markedly higher local curvature than those toward clean classes, and higher curvature at the endpoint predicts lower transfer to clean models. The second is practical. A per-class transfer deficit, calibrated on clean reference-to-reference transfer, detects backdoored models and their target classes in hard-label settings. It should hold up in the regimes where current detectors are documented to fail: multi-target attacks, more than 30% of classes infected, and diffuse blended triggers.

Write $T^{\text{sus}}_c$ for the fraction of targeted perturbations toward class $c$, crafted on the suspect model, that also reach $c$ on the reference models. Write $T^{\text{ref}}_c$ for the same quantity when the perturbations are crafted on one reference model and tested on the others, using the same attack and query budget. The audit statistic is

$$\Delta_c = T^{\text{ref}}_c - T^{\text{sus}}_c ,$$

and a model is flagged when $\max_c \Delta_c$ exceeds a threshold calibrated on clean suspect models of held-out architectures. We test the mechanistic claim as a conditional prediction: after controlling for perturbation norm, the endpoint curvature of a perturbation should predict whether it transfers.

Why This Is Not Already Done

The ingredients exist separately.

Evidence that backdoors are geometrically sharp. Three lines of work motivate the curvature half of the claim. AEVA finds that hard-label adversarial perturbation maps on backdoored models spike sharply at the trigger location, the "adversarial singularity." Model X-ray finds that 2D slices of backdoored boundaries show the target class forming enclosing islands. BURN finds that poisoned samples sit unusually far from the boundary and revert to their true class once pushed across it. None of these papers measures curvature, and none asks whether these distortions transfer.

Evidence that sharpness blocks transfer. On the transfer side, PGN shows that adversarial examples sitting on flat loss maxima transfer better, and it supplies the finite-difference trick we reuse to estimate curvature cheaply. Decision Boundary Adaptation flattens the surrogate model's boundary in parameter space to improve transfer. A Theory of Transfer-Based Black-Box Attacks proves that transfer succeeds when two models share the geometry concentrated around the data. That theory assumes benign training. We read a backdoor as a model-specific addition to this geometry that other models cannot share.

Evidence connecting backdoors and transfer. NTD uses trigger non-transferability as an empirical fact and offers no geometric account. It is also a runtime filter that needs triggered inputs to arrive. Our audit inspects the model before any trigger is seen. Shared Adversarial Unlearning shows that perturbations shared between a poisoned model and a model being repaired bound backdoor risk. That is transfer toward a model that still carries the backdoor. We study transfer toward models that never had it, where the shortcut should be absent.

To our knowledge, no paper in the file measures curvature at backdoor-riding perturbations or uses cross-model transfer failure as a model-level detection statistic.

Research Program

Curvature and transfer measurements. We first build a measurement pipeline on poisoned and clean models from BackdoorBench (codebase).

  • Finding perturbations. For each model and each class we run a targeted minimal-perturbation attack from a fixed pool of clean inputs. There are two access regimes. In the white-box regime the attack is PGD with a norm schedule. In the hard-label regime it is a decision-based attack in the GeoDA/TtBA family (TtBA), which estimates the boundary's normal direction from label queries.
  • Curvature, white-box. At each endpoint we estimate curvature along the boundary with PGN's finite-difference estimate: the change in the margin gradient between two nearby points, divided by the step size.
  • Curvature, hard-label. Here we estimate the normal at several nearby boundary points. Curvature is the rate at which the estimated normal turns per unit distance. This follows the observation of Liu et al. that boundaries near natural data are nearly flat, so normals estimated from few queries are reliable there, while a sharp pocket shows up as fast-turning normals.
  • Curvature sign. Following Fawzi et al., who found that natural points sit on the concave side of the boundary, we also record which side of the boundary each endpoint's curvature bends toward.

Transfer and causal tests. We compute transfer of every perturbation to three to five clean reference models. The references are trained by the defender on the same task, with architectures that differ from the suspect's. We then test the mechanism with two interventions.

  • Flattening the search. We rerun the suspect-model search with PGN's gradient-norm penalty. If backdoor-riding perturbations are sharp, the penalty should move them off the shortcut. The found perturbations should then look less like the trigger, and their transfer to references should rise toward $T^{\text{ref}}_c$.
  • Flattening the model. We apply DBA-style parameter-space flattening to the backdoored model and check whether the backdoor's attack success rate falls. This tests whether the shortcut depends on sharpness, and it is the model-side version of Gap 4 in the file.

The audit. We compute $\Delta_c$ under a fixed query budget per class and calibrate the threshold on clean suspects whose architectures are held out from calibration.

Adaptive attacker. Finally, we test an attacker who knows the defense. The attacker trains the backdoor with a flatness penalty on the trigger region, or aligns the trigger with directions that clean surrogates are already vulnerable to, so that trigger-like perturbations transfer. The account predicts a trade-off. A trigger flat enough to transfer should become a near-universal adversarial perturbation. It should then lose stealth, raising clean-model misclassification, or lose attack success rate. Measuring this trade-off is part of the contribution, not an afterthought.

Baselines

Do nothing. The naive reference point is a detector that flags classes by raw per-class minimal perturbation norm on the suspect model. A backdoor target should be reachable with an unusually small perturbation. This is the entrenchment signal of BURN seen from the other side, and it is cheap. If $\Delta_c$ does not beat it, transfer adds nothing.

AEVA. AEVA is the strongest hard-label competitor under the same threat model. It estimates perturbation maps from labels, extracts per-class peaks, and flags outliers with extreme value statistics. It reports CIFAR-10 detection accuracy of 91.5% to 94.4% and overall accuracy of at least 86.7% using 32 samples per class. We give it the same per-class sample count and a matched query budget.

Model X-ray. Model X-ray is the strongest boundary-geometry detector. It reports AUROC of nearly 1.000 against BadNets and SSBA and 100% detection accuracy on CIFAR-100, using 60 clean samples. We give it 60 clean samples as well. The file reports no numbers for either method under multi-target attacks or with more than 30% of classes infected ("no reported result found on Lacuna" for those settings), so we will produce them.

Fairness conditions. Our method uses reference models, which the baselines do not. We therefore report compute and data cost explicitly, and we also give AEVA and Model X-ray access to the reference models. For AEVA this means calibrating its outlier test on the references' peak distributions instead of across classes. This removes any "extra resource" advantage.

Ablations. Two ablations isolate the key ingredients. The first uses $\Delta_c$ without calibration on reference-to-reference transfer. The second replaces transfer entirely with endpoint curvature, which tests whether the geometry alone detects backdoors.

What counts as a win. Beating the baselines means matching them within noise on single-target BadNets and SSBA, and exceeding them in AUROC on multi-target attacks, attacks infecting more than 30% of classes, and Blended attacks.

Why not just…?

Why not just use AEVA or Model X-ray? Both are excellent on single-target patch triggers, and we do not expect to beat them there. Their documented failures are structural. AEVA's outlier test breaks when more than 30% of classes are infected. Model X-ray's entropy metric weakens under multi-target and Blended attacks. Both failures come from comparing classes to each other. $\Delta_c$ compares each class to clean references, so the multi-target and heavily infected experiments test this advantage directly.

Why not just use NTD-style clean feature comparison? NTD needs a triggered input to arrive at test time. It also degrades on classes with high internal variety, such as the breeds in Animals_10. It answers "is this input triggered," not "is this model compromised." Our audit shares NTD's assumption of a trusted clean model, and we state that assumption plainly. In return it works before deployment and without any triggered inputs. We include Animals_10-like high-variety classes (Tiny ImageNet) to check whether transfer statistics suffer the same overlap problem.

Why not just measure perturbation norm, and forget curvature and transfer? This is the real threat. If backdoor targets are simply closer, norm alone may detect them. It would fail on blended triggers, whose norm budget can resemble ordinary adversarial perturbations. The "do nothing" baseline settles the question. The mechanistic analysis controls for norm when asking whether curvature predicts transfer. If curvature explains nothing beyond norm, we will report the curvature account as refuted and keep $\Delta_c$ only if it beats norm on its own.

Why not just skip curvature and use the transfer deficit as a black-box statistic? For detection, this may be enough, and the curvature-only ablation will show it. Curvature earns its place as the mechanism. It predicts that the PGN penalty moves searches off the shortcut, and it predicts the adaptive attacker's trade-off. A purely empirical transfer statistic makes neither prediction.

Proposed Experiments

Setup. The main grid uses BackdoorBench attacks (BadNets, Blended, WaNet, SSBA) on CIFAR-10, CIFAR-100, GTSRB and Tiny ImageNet. Suspect architectures are PreAct-ResNet18, VGG-19, EfficientNet-B3, MobileNetV3-Large and DenseNet-161. References are drawn from the remaining architectures. We extend the grid with multi-target (all-to-all) variants and with runs infecting 10%, 30% and 50% of classes. The metrics are:

  • AUROC for model-level detection;
  • target-class identification accuracy;
  • queries per class;
  • wall-clock time, compared with Model X-ray's reported 36 seconds on CIFAR-10;
  • for the mechanism, the correlation between endpoint curvature and transfer after controlling for norm, and the change in $T^{\text{sus}}_c$ under the PGN penalty.

Decisive checks. Three results would decide the project.

  1. Curvature gap. Endpoint curvature toward target classes should clearly exceed curvature toward clean classes on the same poisoned model. It should also predict non-transfer beyond norm. If it does not, the geometric account fails.
  2. Detection where others fail. $\Delta_c$ should keep high AUROC in the multi-target and 30%+-infection regimes where the baselines' designs predict failure. If it does not, the audit adds nothing practical.
  3. Adaptive attack trade-off. The flatness-trained backdoor should buy transferability only at a measurable cost in clean-model stealth or attack success rate. If attackers can make backdoors transfer for free, trigger non-transferability is not a law but a default. That would undercut both NTD and this audit, and would be an important negative result in its own right.