When labels are expensive and the data are skewed, active learning picks which unlabeled examples to send to annotators. The usual rule is to pick the examples the model is most uncertain about. When the goal is a model that also resists small adversarial input perturbations, that rule may be the wrong one. We propose a different score: how much the model's prediction on an unlabeled example changes when that input is nudged by a small, worst-case perturbation. We call this fragility. The bet is that fragility finds rare classes better than ordinary uncertainty does, and ignores majority examples that are ambiguous for innocent reasons.
The reasoning comes from two bodies of work that have not been connected. Long-tailed adversarial training shows that rare ("tail") classes pay most of the robustness cost. Adversarial Robustness Under Long-Tailed Distribution found that tail classes have a much larger gap between clean and robust accuracy. Long-tailed Adversarial Training with Self-Distillation found that standard adversarial training can leave tail classes less robust than natural training would. Uncertainty-driven class balancing has the opposite problem: entropy sampling enriches rare classes, but it gets pulled toward noisy, genuinely ambiguous majority examples. If fragility concentrates on tail classes, one attack step per candidate should give a cheaper and more tail-selective acquisition signal than entropy or Monte Carlo dropout. That signal would be aimed at the quantity we care about, which is worst-class robust accuracy per label spent.
Core Claim
We define fragility of an unlabeled input $x$ under the current adversarially trained model $p_\theta$ as
$$F(x) = \mathrm{KL}\big(p_\theta(\cdot \mid x)\,\|\,p_\theta(\cdot \mid x+\delta^*)\big), \quad \delta^* = \arg\max_{\|\delta\|_\infty \le \epsilon} \mathrm{KL}\big(p_\theta(\cdot \mid x)\,\|\,p_\theta(\cdot \mid x+\delta)\big).$$
Here $\epsilon$ is the attack budget and $\delta^*$ is found with one to three gradient steps. This score needs no label, so it can be computed on the unlabeled pool.
We make three testable claims about acquiring the top-$F$ examples each round, with the learner trained adversarially. First, fragility acquires tail-class examples at a higher rate than entropy, because tail classes carry the largest boundary error. Second, it is less attracted than entropy to aleatoric ambiguity, meaning examples that are ambiguous in themselves. Such examples have high clean entropy but need not change much under perturbation. Third, it reaches higher worst-class accuracy under AutoAttack at equal label budgets.
The claim is new because no work in the file evaluates active acquisition for training robust models on imbalanced data (Gap 1). The long-tailed adversarial training papers all train on a fixed, fully labeled set.
Relation to Prior Work
Several papers motivate the proposal. Active Learning as Decision Support for EHR Cohort Construction Under Class Imbalance showed that entropy sampling acts as an "enrichment engine" for rare positives across 18 MIMIC-IV tasks, while diversity methods starved the minority class. Addressing Limited Data for Textual Entailment Across Domains made the same implicit-balancing observation for text. Both use clean uncertainty with non-robust learners. We ask whether the enrichment survives, or improves, once the learner and the score are adversarial.
Other papers supply the balancing machinery we compare against or combine with:
- Inverse-frequency weighting of uncertainty, as in Focused active learning for histopathological image classification and Re-weighting Tokens.
- The "boosted entropy" of PriorAL, which additionally depends on a well-aligned foundation model.
- Monte Carlo dropout mutual information from Prior-free Balanced Replay, which costs many forward passes per candidate.
Fragility replaces the uncertainty term in all of these rather than competing with the frequency weight.
The learner comes from Revisiting Adversarial Training Under Long-Tailed Distributions. That paper showed that adversarial training with Balanced Softmax Loss (AT-BSL) plus diverse augmentation matches or beats the more complex RoBal method. It also showed that long-tailed adversarial training overfits badly, with robustness peaking early and then falling. This constrains our design: querying high-curvature boundary points could make that overfitting worse. We therefore measure overfitting explicitly instead of assuming it away.
The KL-to-perturbed-prediction form of the score is the TRADES-style consistency term used as a selection signal rather than a loss. Improving Adversarial Training using Vulnerability-Aware Perturbation Budget uses per-sample margins to set perturbation budgets on balanced data. We use a related per-sample sensitivity to decide which examples to label. The file's analysis suggests that margin-based budgets would systematically under-defend tail classes (Gap 4). That tension is the reason to test whether sensitivity reflects class scarcity. Finally, Posterior Mismatch Matters warns that peaked one-hot targets hurt tail robustness. We keep the learner fixed so that any gain can be attributed to acquisition alone.
Research Program
The setting is pool-based active learning. The unlabeled pool is itself long-tailed: CIFAR-10-LT and CIFAR-100-LT at imbalance ratio 50, with a Tiny-ImageNet-LT run if compute allows. We start from a small random seed set. Each round we retrain an AT-BSL learner with augmentation at $\epsilon = 8/255$, score the pool, and acquire a fixed batch.
The analysis has three parts. The first is diagnostic: before any active learning, on a fully trained AT-BSL model, we measure how fragility and entropy each correlate with class frequency. This test is cheap and can falsify the premise early. The second is the acquisition study proper, run across budgets up to the point where performance approaches the full-data model. The third checks mechanism. We inject aleatoric ambiguity into head classes, using pixel blends of two head classes with one of the two labels and, separately, 20% label noise following PriorAL's protocol. We then track what fraction of each method's acquisitions land on these corrupted examples.
We also check that the score is not an artifact of gradient masking, where the model's gradients stop pointing toward real adversarial directions. We do this by correlating single-step fragility with fragility from 10-step PGD on held-out examples.
Baselines
The do-nothing reference is random acquisition from the long-tailed pool, trained with the identical AT-BSL learner. Any method that fails to beat it on worst-class AutoAttack accuracy has no value.
Plain predictive-entropy sampling is the strongest single-signal competitor. Active Learning as Decision Support for EHR Cohort Construction Under Class Imbalance found it was the only strategy to beat random consistently across all 18 imbalanced tasks. We compute entropy on the same robust learner so that the only difference is the score.
Entropy multiplied by the inverse frequency of the predicted class is the most dangerous competitor. It combines the acquisition rule of Focused active learning with the frequency weighting of PriorAL, without PriorAL's foundation model.
Monte Carlo dropout mutual information, following Prior-free Balanced Replay, represents epistemic uncertainty estimation. Coreset represents diversity selection.
The full-data AT-BSL model gives the ceiling. For reference, the strongest reported full-data worst-class robust accuracy on CIFAR-10-LT at IR=50 is 26.23% for PAT, against about 15.59% for AT-BSL, as reported in Posterior Mismatch Matters. RoBal reports 32.04% average AutoAttack accuracy in Adversarial Robustness Under Long-Tailed Distribution. For active acquisition with a robust learner on long-tailed pools, no reported result was found on Lacuna, so all active baselines are run by us.
Fairness conditions are as follows. Every method uses:
- the same ResNet-18 learner, seeds, round schedule, and batch size;
- the same retraining recipe, with early stopping on a small balanced validation split;
- a matched hyperparameter search over acquisition batch size.
We report acquisition cost in forward-backward passes per candidate. Fragility with one step costs about two, while Monte Carlo dropout with $T$ samples costs $T$.
Two ablations isolate the key ingredient. The first computes fragility on a naturally trained learner, to test whether the robust model is what makes the score tail-selective. The second multiplies fragility by inverse predicted-class frequency, to test whether it complements explicit balancing.
Beating the baselines means reaching the same worst-class AutoAttack accuracy with at least 25% fewer labels than weighted entropy, averaged over three seeds with non-overlapping confidence intervals, without losing more than one point of clean accuracy.
Why not just…?
Why not just use entropy sampling, which already enriches rare classes? It might be enough. The EHR evidence is strong, but it concerns clean, low-dimensional clinical embeddings and non-robust learners. The file also notes that entropy breaks down when majority classes contain aleatoric noise. The injected-ambiguity experiment tests exactly this. If entropy acquires no more corrupted head examples than fragility does, and matches it on worst-class robust accuracy, then the fragility claim fails.
Why not just multiply entropy by inverse class frequency? This is the most serious threat, because explicit reweighting may capture most of the tail enrichment. It still ranks examples within a class by clean ambiguity, which rewards noisy examples. It also depends on predicted labels, which are unreliable for the tail early in acquisition. We run it as a first-class baseline and also combine it with fragility. If the combination is no better than weighted entropy, fragility adds nothing, and we will report that.
Why not skip active learning and run AT-BSL with augmentation, or self-distillation, on a random labeled subset of the same size? These are strong full-data methods: AT-BSL with augmentation gained 6.66% AutoAttack accuracy over RoBal in Revisiting Adversarial Training Under Long-Tailed Distributions. But they cannot create tail examples that were never labeled. The random-acquisition baseline trained with AT-BSL is exactly this alternative. The per-class curves will show whether the remaining gap is caused by missing tail data, which acquisition can fix, or by the training objective, which it cannot.
Decisive Checks
The first check is correlational. On a trained AT-BSL model for CIFAR-10-LT, if the rank correlation between per-class mean fragility and class frequency is not clearly more negative than the same correlation for entropy, the mechanism is absent and the program stops there.
The second check is the label-efficiency curve for worst-class AutoAttack accuracy against weighted entropy, together with the tail fraction of acquired labels.
The third check is the robust-overfitting gap, measured as best-epoch minus last-epoch robust accuracy for learners trained on each method's acquired set. If fragility-acquired sets overfit markedly more, the resolution of the Gap 1 tension is negative. That would be a useful finding in itself, and it would point toward pairing fragility with the augmentation remedy.
All robust numbers are reported under AutoAttack, not PGD alone, because of the "fake gain" failure documented in Adversarial Robustness Under Long-Tailed Distribution, where PGD overstated robustness on long-tailed data.
References
- Adversarial Robustness Under Long-Tailed Distribution
- Long-tailed Adversarial Training with Self-Distillation
- Active Learning as Decision Support for EHR Cohort Construction Under Class Imbalance
- Addressing Limited Data for Textual Entailment Across Domains
- Focused active learning for histopathological image classification
- Re-weighting Tokens: A Simple and Effective Active Learning Strategy for Named Entity Recognition
- Active Learning with Foundation Model Priors: Efficient Learning under Class Imbalance
- Prior-free Balanced Replay: Uncertainty-guided Reservoir Sampling for Long-Tailed Continual Learning
- Revisiting Adversarial Training under Long-Tailed Distributions
- Improving Adversarial Training using Vulnerability-Aware Perturbation Budget
- Posterior Mismatch Matters: Adversarial Training for Long-Tailed Robustness