Do Transfer Attacks Ride on Augmentation Fingerprints?

Adversarial examples crafted against one image classifier often fool a different classifier that the attacker has never queried. This property, adversarial transferability, is what makes zero-query attacks on private models possible. The usual explanation is that models trained on the same task learn similar features. We propose a narrower hypothesis that can be tested directly. Models trained with the same data-augmentation recipe (the random crops, flips, and color changes applied to training images) may also share internal directions that record the augmentation itself: where the crop was taken, whether the image was flipped, how strongly colors were jittered. The strongest transfer attacks, which perturb a surrogate model's intermediate features, may partly succeed by pushing along these shared "augmentation fingerprint" directions rather than along semantic ones.

To vary how much augmentation information sits at each depth, we use local learning. In local learning a network is split into blocks, and each block is trained by its own small objective with no gradient arriving from later layers. Without that top-down signal, an early block has nothing that tells it to discard nuisance information. Local training should therefore change how much augmentation information survives at a given depth. We compare local and end-to-end models, measure that information layer by layer, and then surgically remove it from the attack objective. This tells us whether augmentation fingerprints cause transferability or merely correlate with it.

Core Claim

We claim that part of intermediate-level transferability is carried by feature directions that encode augmentation parameters. The claim makes three predictions. First, attacks restricted to the augmentation-predictive subspace of a surrogate layer should transfer well to targets trained with the same augmentation pipeline and poorly to targets trained with a different one. Second, projecting that subspace out of the attack objective should reduce transfer to pipeline-matched targets more than projecting out a random subspace of equal dimension. Third, the effect should scale with how much augmentation information a layer holds. That amount should differ between greedily trained local blocks and end-to-end layers at the same depth.

The prediction for contrastive local learning goes the other way, and it is a useful internal check. Preventing Dimensional Collapse in Contrastive Local Learning with Subsampling shows that strong augmentations drive early decoupled contrastive blocks into dimensional collapse, meaning their representations use only a few effective directions. A contrastive objective explicitly rewards invariance to augmentation. We therefore expect surviving dimensions in contrastive local blocks to carry less augmentation information, while supervised local blocks with classification heads carry more. If both kinds of local block behave identically, our account of why local training matters is wrong.

Relation to Prior Work

The work closest to ours is ILA-DA. It maximizes disruption of a surrogate's intermediate features, using learned input augmentations, a "super-clean" reference image, and interpolation between attack iterates. ILA-DA uses augmentation as a tool applied to the attack's inputs. We ask what augmentation left behind inside the models' features, and whether attacks exploit it. ILA-DA also picks the attack layer heuristically and uses only end-to-end surrogates. We replace the heuristic with a measured per-layer profile and add locally trained surrogates.

Two papers motivate the hypothesis. Publishing Efficient On-device Models Increases Adversarial Vulnerability shows that shared training lineage aligns decision geometry enough to enable transfer. It studies pruning and quantization, not training recipe. Adversarial Vulnerability from Interference Between Features in Superposition ties vulnerability to interference between features packed into too few dimensions. Augmentation-encoding directions are a concrete candidate for such interfering features.

The augmentation-leakage literature supplies the intuition but lives in generative models. Improving the Leaking of Augmentations in Data-Efficient GANs via Adaptive Negative Data Augmentation shows that generators reproduce augmentation artifacts even under supposedly leak-free pipelines. Towards Bridging the Performance Gaps of Joint Energy-Based Models shows that hybrid models trained on augmented data inherit padding and cropping artifacts. We carry the notion of leakage into discriminative representations, where the artifact is information in a feature rather than a visible defect in a sample.

The local-learning literature supplies the knob. Go beyond End-to-End Training: Boosting Greedy Local Learning with Context Supply models greedy local training as a chain in which each block decides, irreversibly, what to keep. LoCo and Momentum Auxiliary Network for Supervised Local Learning restore some downstream feedback, through overlapping blocks and momentum-shared context respectively. Together these give us a range from fully greedy training to near end-to-end, holding the architecture fixed.

Neither LoCo nor MAN has ever been evaluated adversarially. The only local-learning adversarial study in the file, Toward Spiking Neural Network Local Learning Modules Resistant to Adversarial Attacks, concerns spiking networks and not transfer.

Research Program

Surrogate and target zoo. On CIFAR-10 we train ResNet-18 and VGG-style targets under four augmentation pipelines: none, crop-and-flip, crop-and-flip plus cutout, and SimCLR-strength color jitter plus crops. We train surrogates of the same architectures in five ways:

  • end-to-end supervised;
  • greedy supervised local learning with linear heads at 4, 8, and 16 blocks;
  • MAN-style local learning;
  • decoupled SimCLR;
  • subsampled decoupled SimCLR, following Fournier et al.

Every surrogate is also trained under each pipeline. This lets us cross surrogate pipeline with target pipeline while holding the training method fixed.

Leakage profile. For each layer we measure how much augmentation information is present. We fit linear probes on frozen features of augmented inputs to predict crop offset and scale (binned), flip, and jitter strength. The profile is probe accuracy, or R² for continuous parameters, above a chance-level reference from the same layer of a model trained without that augmentation. We record class-probe accuracy alongside, so that augmentation information and semantic information are separated.

Attack decomposition. At each layer we estimate the augmentation subspace as the span of the probe weights. We run the ILA objective in three versions: unrestricted, with the subspace projected out ("semantic-only"), and inside the subspace only ("nuisance-only"). As a control, we also project out a random subspace of matched dimension.

ImageNet check. Finally, we test the most informative configurations on ImageNet with an end-to-end ResNet-50 surrogate. We use the ILA-DA protocol and publicly available targets whose training recipes differ, to check that the effect is not a small-image artifact.

Baselines

Naive reference: I-FGSM. The do-nothing reference is I-FGSM, an iterative gradient attack on the surrogate's output loss. It is run with the same perturbation budget and iteration count as every other attack. On ImageNet, ILA-DA reports I-FGSM at 7.3% transfer success against ensemble adversarially trained models (ILA-DA).

Standard intermediate attack: ILA. Plain ILA uses the layer chosen by the original heuristic. It isolates what intermediate-level attacking contributes before any subspace surgery.

Strongest competitor: ILA-DA. ILA-DA is the strongest existing method and is exactly what our mechanism claims to partly explain. Its reported ImageNet results are 84.5% average transfer success across nine undefended black-box models, 75.9% against ensemble adversarially trained models, and 56.3% from CNNs to Vision Transformers. For CIFAR-10 we found no reported ILA-DA result on Lacuna, so we rerun it ourselves.

Fairness is essential for this comparison. We run ILA-DA with its own iteration schedule (up to 500 iterations on defended targets). Our subspace variants get the same iteration budget and the same augmentation search. We are not trying to beat ILA-DA's headline number. "Beating" ILA-DA means showing that a nuisance-only variant recovers a large fraction of its success on pipeline-matched targets, and that semantic-only ILA-DA loses clearly more transfer on matched targets than on mismatched ones.

Key ablation: random projection. The random-subspace projection isolates our key ingredient. If removing a random subspace of the same size hurts transfer as much as removing the augmentation subspace, the hypothesis fails.

Local-learning reference points. For clean accuracy we anchor to Fournier et al.'s CIFAR-10 results at 16 blocks: 92.1% end-to-end, 85.9% decoupled SimCLR, and 87.1% subsampled. Our local surrogates must land near these before their transfer numbers mean anything.

Why not just…?

Why not just vary the target's augmentation pipeline and measure transfer, without probes or local learning? We do run this, and it is the first result we report. It cannot separate mechanisms, though. Pipeline-matched models may transfer better because they learned more similar semantic features, not because they share augmentation fingerprints. The nuisance-only and semantic-only attacks, compared against the random-projection control, are what distinguish the two stories. If pipeline matching has no effect on transfer at all, the project stops there, and that negative result is itself useful.

Why not just use end-to-end surrogates, since ILA-DA already works so well on them? This is a real threat to the local-learning part of the proposal. If end-to-end surrogates show the same relationship between leakage and transfer across layers, local learning serves only as a convenient way to vary leakage and is not a finding in itself. We will report it that way. The contrast between supervised and contrastive local blocks is the test that local training does something end-to-end training cannot. It is also practically relevant: memory-saving local training is a natural fit for the edge models that Hong et al. show become attack surrogates.

Why not just select the attack layer by class-probe accuracy? Choosing the layer by how well it separates classes is a simpler rule than ours. We fit both predictors, class information and augmentation information, against per-layer transfer success. If augmentation information adds no explanatory power once class information is controlled for, our layer-selection claim is unnecessary.

Decisive Checks

Metrics are untargeted transfer success rate on each target, broken down by whether the target's pipeline matches the surrogate's, plus per-layer probe scores. The hypothesis survives only if all four of the following hold:

  1. Nuisance-only attacks transfer noticeably better to pipeline-matched targets than to mismatched ones.
  2. Removing the augmentation subspace lowers matched-target transfer more than random removal does.
  3. Across layers and surrogates, augmentation information predicts transfer to matched targets beyond what class information predicts.
  4. Supervised local blocks show more augmentation information than end-to-end layers at equal depth, while contrastive local blocks show less.

A positive outcome has a defensive corollary that we test last. Training a target with a non-standard augmentation pipeline should reduce transfer from publicly available surrogates at almost no cost to clean accuracy.