Poison Filtering That Spares the Tail: Backdoor-Robust Data Pruning on Imbalanced Data

Two ideas meet here. The first is data pruning: shrinking a training set by keeping only a well-chosen subset of examples, called a coreset. The second is the backdoor attack: an adversary slips a small number of "poisoned" training examples into the data. Each carries a hidden trigger, such as a small patch or an invisible image warp, and is relabeled to a target class. The trained model then behaves normally except when the trigger appears.

The best recent backdoor defenses are themselves pruning methods. They discard examples the network learns suspiciously fast, because triggers are simple shortcuts. The pruning literature has shown the flip side of this logic. Scoring examples on a single global scale mistakes rare-class examples for noise and deletes them. On long-tailed data, where a few classes are common and most are rare, "how fast was this learned" means something different for each class. Current defenses ignore this.

We propose to judge each example's learning speed against what its class size predicts, rather than against a global threshold. A clean rare-class example should be learned slowly. A cluster of examples labeled with a rare class that is learned as fast as a head class is the signature of a trigger. We use this calibrated signal only to decide which examples to remove and how many to keep per class. The survivors inside each class are then sampled at random. The claim is that this pipeline removes backdoors on long-tailed data while keeping tail-class accuracy. It should succeed exactly where both global and per-class-ranked entropy filtering fail: when the attacker targets a rare class, so that poisons outnumber that class's clean examples.

Why the obvious combinations break

The closest existing work is Anti-Backdoor Coreset Selection via Cumulative Entropy (ABCS). It sums each example's prediction entropy over early epochs, keeps the highest-entropy roughly 55%, and retrains on that subset. It reports over 95% defense effectiveness on CIFAR-10 and attack success near 1%.

Two related defenses also filter globally. COLLIDER keeps examples whose gradients sit near the center of the gradient cloud and that have low local intrinsic dimensionality, a measure of how sparse an example's neighborhood is. Anti-Backdoor Learning (ABL) isolates the lowest-loss 1% of examples early in training and then unlearns them with gradient ascent.

All three were evaluated only on roughly balanced data. On long-tailed data, each global criterion lines up with class frequency rather than with poisoning:

  • ABL's low-loss pool fills with easy head-class examples.
  • COLLIDER's sparse-neighborhood filter targets tail classes, which live in sparse regions by nature.
  • ABCS's high-entropy selection over-keeps ambiguous tail examples and starves head classes.

This mirrors the class deletion that DRoP: Distributionally Robust Data Pruning documents: one global-score baseline "literally deleted entire classes" on CIFAR-100 at 10% density.

The minority-preserving pruners fail in the other direction when data are poisoned. DRoP assigns each class a retention quota that grows with the class's early error rate (its miss rate, one minus recall). If its error estimate is computed on data drawn from the poisoned pool, the target class looks easy, because the trigger is learned almost immediately. DRoP then shrinks that class's quota. Random subsampling inside the class preserves the poison ratio while throwing away clean target-class examples.

The supplied file raises, but does not settle, a concern about SIMS. At aggressive budgets SIMS favors easily separable, prototypical examples, which is exactly how poisons look.

The natural patch is "stratified ABCS": rank by entropy within each class and keep a quota proportional to the class's observed size. That quota is inflated by the poisons, since all poisons carry the target label. When the target is a tail class, poisons can be the majority of that class. Any within-class ranking or normalization then treats poisons as the typical member.

Core Claim

A trigger makes a set of examples learn at the speed of a class the size of the poison set, not the size of the class they are labeled with. That gap can be measured across classes. We write it as a robust fit over classes:

$$ \bar H_c \approx \alpha + \beta \log n_c , $$

In this fit, $\bar H_c$ is the median cumulative entropy of examples labeled class $c$ after a short warm-up, $n_c$ is the observed count of class $c$, and $\alpha,\beta$ are fitted with a robust regression that tolerates a few outlying classes. Under an all-to-one attack, only the target class breaks this relationship. It has both an inflated count and an entropy that is too low, or, in the tail-target case, a distinctly bimodal entropy distribution.

Within a flagged class, we separate the fast, tight component from the slow, diffuse one using a two-component mixture over entropy trajectories. We do not use a rank cutoff, because a mixture still works when the fast component is the majority. We then estimate each class's clean count and clean error from its retained component only. Quotas are set DRoP-style from these cleaned estimates, and the retained pool is subsampled at random inside each class. We keep random within-class sampling because DRoP found that score-based selection inside a class hurts class-conditional generalization. Scores are used only for rejection and quota setting.

What is new relative to ABCS is twofold. First, a cross-class, frequency-calibrated reference for "too easy." Second, the separation of three steps: rejecting poisons, allocating quotas, and choosing within a class. What is new relative to DRoP is quota estimation that is robust to label-inflating poisons.

Research Program

The work has three parts, each testing a claim that could fail.

The first part is a diagnostic study. We run ABCS, COLLIDER, ABL, DRoP and SIMS on poisoned long-tailed data and measure what each actually removes, split by class rank and by poison status. We attack a head target, a middle target and a tail target. This tests Gaps 1–3 of the file directly. The key open question comes from ABL's assumption: does a trigger aimed at a tail class still converge faster than clean head examples? Or does the tail class's low update frequency mask it? We also check whether DRoP really shrinks the target class's quota. If tail-targeted triggers turn out to be masked, so that they are not fast even in absolute terms, the calibration idea loses its premise. In that case we would report this as the main finding.

The second part builds the calibrated selector described above. Its inputs are the warm-up statistics ABCS already computes, so it adds essentially no training cost. We will also test whether $\log n_c$ is the right covariate by fitting the class-level relationship on clean long-tailed runs first. If the relationship is noisy, we replace it with a nonparametric monotone fit.

The third part is adaptive stress. We run blend-style attacks that ABCS already fails to fully neutralize, and a warping attack in the style of WaNet. We also add an attacker who knows our defense. This attacker spreads poisons across several target classes, so that no single class is a regression outlier. It also adds noise-mode-style clean augmentations to broaden the poison component's entropy.

Baselines

The naive reference points are two. The first is training on the full poisoned set with no defense, which bounds attack success from above. The second is random pruning to the same budget. Random pruning shows how much of any gain comes merely from keeping less data, since it dilutes nothing and deletes no class.

The minority-preserving pruners, DRoP and SIMS, are run on the poisoned data exactly as published. They show what a practitioner who cares about fairness but ignores poisoning would get.

The strongest competitor is ABCS. It is the only method that already frames backdoor removal as coreset selection and reports the best numbers on the file's benchmarks: over 95% defense effectiveness on CIFAR-10, attack success near 1%, and training on about 55% of the data. We will run it with the authors' warm-up length, unlearning step and coreset size.

The second-strongest competitor is stratified ABCS, meaning cumulative entropy ranked within each class under label-matched quotas as recommended by Disentangling the Roles of Representation and Selection in Data Pruning. This is the baseline a reviewer would build first, so it must be tuned as carefully as our method.

ABL and COLLIDER complete the defense set. For reference on balanced CIFAR-10:

  • ABL reports average attack success of 7.69% across ten attacks, with 84.76% clean accuracy.
  • COLLIDER reports attack success of 4.80% under BadNets, with 80.66% clean accuracy.

We found no reported result on Lacuna for any backdoor defense on long-tailed or imbalanced data, so all long-tailed comparisons will be new measurements.

Fairness rules apply to every method. All methods share one architecture, the same warm-up epoch budget, the same final retained fraction (swept from 55% down to 10%), and the same total GPU hours. Each gets an equal hyperparameter search over a small untrusted validation split drawn from the poisoned pool. None gets a trusted clean set.

Three ablations isolate the ingredients. The first removes the cross-class calibration, leaving only the within-class mixture. The second replaces random within-class sampling with entropy ranking. The third sets quotas from raw observed counts instead of cleaned estimates.

Beating ABCS means two things at a matched budget. Attack success must be no higher than ABCS's on balanced CIFAR-10, so we give nothing up where it works. On long-tailed data, we must deliver lower attack success and higher tail-class and worst-class accuracy, with the tail-target setting as the headline case.

Why not just…?

Why not just run stratified ABCS? This is the real threat, and we expect it to match us when the target is a head class, because poisons are then a small minority of that class and a within-class rank separates them. Our claim rests on the tail-target regime. There, poisons form the bulk of the target class, so a within-class rank or z-score calls the poisons "typical." Stratified ABCS also inherits the inflated quota, and DRoP's evidence against score-based within-class selection argues against its selection rule. The decisive experiment sweeps the ratio of poisons to clean target-class examples from 0.1 to 10. If stratified ABCS keeps pace across that sweep, the cross-class calibration is unnecessary, and we will say so.

Why not just remove the backdoor after training with Fine-Pruning? Fine-Pruning cuts neurons that are dormant on clean data and then fine-tunes. It needs a verified clean set, which our setting lacks. It left about 30% attack success on US Traffic Signs, and WaNet was shown to bypass it. On long-tailed data, a clean set with enough tail examples to fine-tune on is also exactly what is scarce. We include it only as a reference that is given a clean set.

Why not just run ABL to clean the data, then prune with DRoP? ABL's global low-loss isolation should mostly catch easy head-class examples when classes are imbalanced. Its gradient ascent would then damage head-class accuracy while missing slow tail-target triggers. The diagnostic study measures the poison purity of ABL's isolated pool by class, and we run the ABL-then-DRoP pipeline end to end as a baseline.

Why not just reweight the classes, or use distributionally robust training (which upweights the worst-performing groups), instead of pruning? Neither saves compute, which is the point of pruning. Both also upweight hard examples, which could amplify blend-style poisons that look hard. We run class-balanced reweighting on the full poisoned set as a non-pruning reference.

Proposed Experiments

The datasets are the following:

  • Long-tailed versions of CIFAR-10 and CIFAR-100, built by exponential subsampling with imbalance ratios of 10 and 100.
  • GTSRB, whose class sizes are naturally uneven.
  • iNaturalist, used in SIMS, as a natural long tail at scale.
  • Balanced CIFAR-10 and Tiny-ImageNet, to confirm we do not regress where ABCS was tested.

The attacks are BadNets, Blend, adaptive blend, WaNet, and our adaptive multi-target variant. Poison rates range from 1% to 10%, and targets are head, middle and tail classes.

We report five metrics:

  • clean accuracy;
  • worst-class accuracy and mean accuracy over the rarest third of classes;
  • attack success rate;
  • defense effectiveness rate as ABCS defines it;
  • precision and recall of the discarded set with respect to true poisons.

Two checks are decisive. The first is the tail-target sweep over the poison-to-clean ratio against stratified ABCS. It tests whether cross-class calibration matters, and is the result that could falsify the proposal. The second is the aggressive-budget regime (10–30% retained) on long-tailed CIFAR-100. There, global pruning is known to delete classes, and we must show that we keep every class while holding attack success near the level ABCS reports on balanced data.