Adversarial training is a defense in which the model learns on deliberately worst-case perturbed copies of its inputs. Run over a whole poisoned dataset, it turns out to be one of the more reliable ways to neutralize backdoors, the hidden triggers an attacker plants in training data so that the model obeys a chosen label whenever the trigger appears. It has two costs. Every training step needs several extra forward and backward passes to find the perturbations, and the model loses accuracy on ordinary clean inputs. This proposal argues that both costs are mostly wasted, because a backdoor lives in a small number of training samples. The plan is to apply adversarial training only to the few percent of samples that cheap signals from training itself mark as suspicious, and to train every other sample normally.
The reason this might work comes from the attacker's own constraint, the stealth-effectiveness trade-off. A trigger strong enough to reliably hijack the model also tends to be learned unusually fast and to dominate the samples it sits in, which makes those samples easy to flag. A trigger subtle enough to avoid being flagged is weak, and weak triggers are the ones a small adversarial perturbation budget erases. If that holds, the attacker is caught either way. Selective adversarial training should then match the backdoor suppression of full adversarial training while recovering most of the clean accuracy and most of the compute.
Core Claim
Our claim is that a per-sample, suspicion-gated version of adversarial training can reach attack success rates comparable to full Composite Adversarial Training on standard backdoor benchmarks. We expect it to keep clean accuracy within about a point of an undefended model and to cost roughly $1+\rho M$ times standard training rather than $M+1$ times, where $\rho$ is the fraction of samples given adversarial treatment and $M$ is the number of inner attack steps. Concretely, we train
$$\min_\theta \sum_{i=1}^{n} \Big[(1-s_i)\,\mathcal{L}(f_\theta(x_i),y_i) + s_i \max_{\|\delta\|\le \epsilon_i}\mathcal{L}(f_\theta(x_i+\delta),y_i)\Big],\qquad \sum_i s_i \le \rho n,$$
where $f_\theta$ is the classifier with weights $\theta$, $\mathcal{L}$ is cross-entropy, and $s_i\in\{0,1\}$ marks whether sample $(x_i,y_i)$ gets adversarial treatment in the current epoch. The indicator $s_i$ is recomputed each epoch from a suspicion score. Each selected sample has its own perturbation radius $\epsilon_i$, which grows with its suspicion.
What would be new is the claim itself, together with an empirical map of the attacker's frontier under this defense. That map measures how much attack success an adaptive attacker can keep while staying below the selection threshold.
Relation to Prior Work
Three papers motivate the proposal. On the Effectiveness of Adversarial Training Against Backdoor Attacks applies adversarial training under several perturbation types as a backdoor defense. It shows that the Composite variant (CAT) drives every tested attack's success rate below 6% on CIFAR-10, and that perturbations bounded in max-pixel ($L_\infty$) norm handle whole-image triggers such as Blended and WaNet. It perturbs every sample for the entire run and does not address the clean-accuracy drop. It also reports that raising $\epsilon$ can at first increase attack success, which is a warning that we take seriously when choosing $\epsilon_i$. Anti-Backdoor Learning (ABL) exploits the fact that poisoned samples are learned unusually fast: it isolates the lowest-loss ~1% of samples and applies gradient ascent to them. ABL fails against naturalistic triggers (SIG, Blend) that are not learned fast. Not All Poisons are Created Equal (EPIC) finds that only about 8% of injected poisons drive a targeted poisoning attack and removes them cheaply as gradient outliers. Together these three papers suggest that the useful part of a backdoor defense is concentrated in a small set of samples. We differ from all three in method: we do not discard or ascend on suspicious samples. We harden them adversarially, which still helps when the selection is imperfect.
The closest mechanism is TART, which also chooses per sample between full $\epsilon$ and $\epsilon=0$. TART makes that choice from the geometry of the data manifold, estimated with an autoencoder and local PCA, and it does so to raise clean accuracy against test-time attacks. It never considers poisoning. Its per-sample gate reaches 83.73% clean accuracy versus 81.76% for standard adversarial training, at 1.24x the cost of standard adversarial training. That result is evidence that gating does not destroy robustness. Our gate is driven by signals that indicate poisoning rather than by geometry, and it is meant to cut cost relative to standard training rather than add to it.
Two papers constrain the design. Rethinking Backdoor Attacks argues that an effective trigger is the strongest feature in its samples and not a statistical outlier, so outlier clustering fails. This supports scoring samples by how strongly the model depends on them, and argues against cluster-based scoring. Backdoor Attacks in Token Selection of Attention Mechanism proves, for a single-head attention model, that a trigger must sit in a "Goldilocks zone" of strength. Below that zone the trigger is not learned; above it, clean accuracy collapses. This is the theoretical form of the trade-off our defense depends on.
On the attack side, Few-shot Backdoor Attacks via Neural Tangent Kernels (NTBA) and the forgetting-based sample selection of Data-Efficient Backdoor Attacks provide the strongest low-footprint attackers we test against. NTBA reaches 90% attack success with 30 images.
Research Program
The suspicion score uses only quantities that ordinary training already produces or that cost one extra backward pass. The first signal is how fast a sample's loss falls relative to others with the same label over a sliding window of epochs, which is the signal ABL relies on. The second is how concentrated the gradient of the loss with respect to the input is: the share of that gradient falling on the top few percent of pixels or frequency bands. A sample whose label is carried by a small, strong pattern should score high on this signal even if it is learned slowly, which targets the samples ABL misses. The two signals are rank-combined within each class. The top $\rho n$ samples get adversarial treatment with $M$-step projected gradient attacks, and their $\epsilon_i$ is interpolated between a floor and a ceiling according to rank. A small random slice of the adversarial budget, about 10% of it, goes to unflagged samples so that no sample is permanently exempt. The schedule starts with a short warm-up of plain training so that the loss-speed signal exists, which mirrors ABL's isolation phase.
The analysis has two parts. First we measure precision and recall of the selection against ground-truth poison labels over the course of training. Then we run adaptive attackers who know the defense and optimize triggers to stay below the selection threshold, by penalizing fast loss decrease and concentrated input gradients during trigger design. For each attacker we trace attack success against poison ratio and trigger strength. The proposal's central bet is that this frontier bends down: evading selection should force either many more poisons or a much lower attack success rate.
Baselines
The naive reference is undefended standard training on the poisoned data. It gives the upper bound on clean accuracy and on attack success, and its compute defines the 1x cost unit. We will run it through the BackdoorBench codebase with that benchmark's default recipes.
The strongest competitor on our exact question is full Composite Adversarial Training from On the Effectiveness of Adversarial Training Against Backdoor Attacks. It is the method we claim to match on attack suppression while beating on cost and clean accuracy. The cited paper reports that it reduced all tested backdoor attacks' success rates below 6% on CIFAR-10, and that spatial adversarial training kept approximately 85% clean accuracy against BadNets and LC. The file contains no reported wall-clock cost for CAT and no CAT clean accuracy on CIFAR-100, so we will measure both. To keep the comparison fair, CAT and our method use the same architecture, the same data and poisons, the same inner-step count $M$, and the same $\epsilon$ ceiling, and both get the same hyperparameter search budget. We beat CAT if our attack success stays within 2 points of CAT's on every attack while we use under a third of its GPU-hours and gain clean accuracy.
The second competitor is ABL. It reports 7.69% average attack success across 10 attacks on CIFAR-10 with 84.76% clean accuracy, against 88.41% for a model trained on clean data. ABL is the cheapest in-training defense, and its isolation step is close to our first suspicion signal. We will run it in BackdoorBench with its published isolation ratio. Beating ABL means lower attack success on SIG and Blend, where it is known to fail, at no loss of clean accuracy.
We also include Augmented Neural Fine-Tuning (NFT), a post-training soft-masking method that reports attack success falling from nearly 100% to 1.81% on CIFAR-10, with clean accuracy within 1% of the original, in 28.4 seconds. NFT assumes a trusted clean subset, which our setting does not. We will report it both with the clean subset and with that subset drawn from the poisoned pool, so the reader can see what the assumption buys.
Three ablations isolate the key ingredient. The first gives the same adversarial budget $\rho$ to uniformly random samples. The second uses only the loss-speed signal. The third uses only the gradient-concentration signal. A fourth control drops flagged samples instead of hardening them, in the style of EPIC, which tests whether adversarial treatment matters beyond simply removing data.
Why not just…?
Why not just run adversarial training on a random subset of the same size? This is the most serious threat, and the random-budget ablation tests it directly. Our expectation that it fails rests on mechanism. Adversarial training stops the trigger-to-label shortcut only when the poisoned samples themselves are perturbed, and a random 10% subset perturbs only about 10% of the poisons. If random allocation matches targeted allocation, the selection signals contribute nothing. In that case we would report that result and reframe the work as a cost study of subsampled CAT.
Why not just use single-step adversarial training on everything? It cuts cost, but the file notes that single-step and approximate inner loops (Reducing Adversarial Training Cost with Gradient Approximation) suffer catastrophic overfitting. It also still pays the clean-accuracy penalty on every sample. We include a single-step-everywhere arm at matched GPU-hours. We expect it to cost more clean accuracy than our method, and we will check whether its suppression of whole-image triggers holds up.
Why not just isolate and discard suspicious samples, as ABL or EPIC do? Hard removal depends on precision at the top of the ranking. ABL's documented failure on SIG and Blend shows that naturalistic poisons are not ranked at the top. Removal also deletes legitimate rare samples, the over-cleansing problem documented in Training set cleansing of backdoor poisoning by self-supervised representation learning. Hardening degrades gracefully: a mis-flagged clean sample receives a perturbed but correctly labeled update rather than being deleted. The drop-versus-harden control settles this.
Why not just augment heavily or prune afterwards? Color Backdoor reports that aggressive augmentation destabilizes training and lowers clean accuracy. Spatial transformations miss whole-image triggers, as the CAT paper shows. Pruning methods such as ANP damage features that clean inputs also rely on, and in NFT's one-shot comparison competing methods left attack success above 86%. We include Fine-Pruning from BackdoorBench as a reference point but do not expect it to compete.
Proposed Experiments
The main grid follows BackdoorBench on CIFAR-10, CIFAR-100, GTSRB, and Tiny ImageNet, with PreAct-ResNet18 as the primary model and EfficientNet-B3 as a check across architectures, since BackdoorBench shows that attack success varies sharply between architectures. Attacks are BadNets, Blended, WaNet, SIG, and label-consistent attacks at poisoning ratios from 0.5% to 10%. We add two low-footprint attackers: NTBA, at the image subsets and activations it supports, and forgetting-based sample selection. Metrics are clean accuracy, attack success rate, and robust accuracy in BackdoorBench's sense (the share of poisoned inputs assigned their true class), which BackdoorBench shows collapses as the number of classes grows. We also report total GPU-hours, the precision and recall of the selection step, and, as a secondary result, AutoAttack accuracy on clean test data.
We sweep $\rho$ over 2%, 5%, 10%, and 20%, and sweep the $\epsilon$ ceiling, to trace cost against attack success against clean accuracy. These curves are plotted against the single operating points of CAT and ABL.
There are three decisive checks. The first is the random-budget ablation at equal $\rho$: targeted selection must cut attack success substantially relative to random selection, or the core claim fails. The second is the SIG and Blend subset, where the gradient-concentration signal must rescue cases that ABL-style loss speed misses; we will report whether it does. The third is the adaptive-attacker frontier. If an attacker can keep more than 50% attack success at the poisoning ratios BackdoorBench uses while staying unflagged, then the stealth-effectiveness trade-off does not protect this defense, and we will say so.
References
- On the Effectiveness of Adversarial Training against Backdoor Attacks
- Anti-Backdoor Learning: Training Clean Models on Poisoned Data
- Not All Poisons are Created Equal: Robust Training against Data Poisoning
- TART: Boosting Clean Accuracy Through Tangent Direction Guided Adversarial Training
- Rethinking Backdoor Attacks
- Backdoor Attacks in Token Selection of Attention Mechanism
- Few-shot Backdoor Attacks via Neural Tangent Kernels
- Data-Efficient Backdoor Attacks
- BackdoorBench: A Comprehensive Benchmark of Backdoor Learning
- Augmented Neural Fine-Tuning for Efficient Backdoor Purification
- Reducing Adversarial Training Cost with Gradient Approximation
- Training Set Cleansing of Backdoor Poisoning by Self-Supervised Representation Learning
- Color Backdoor: A Robust Poisoning Attack in Color Space
- Adversarial Neuron Pruning Purifies Backdoored Deep Models