Backdoor attacks secretly train an image classifier to output an attacker's chosen label whenever a small pattern, the "trigger," appears in the input. Recent attacks get much stronger by being picky about which training images they poison. Instead of choosing images at random, they choose hard ones: images the model is unsure about, near its decision boundary. On such images the clean evidence is too weak to fit the label, so the model reaches for the trigger as a shortcut. The same choice also hides the attack from standard defenses that search for poisoned images clustering apart from clean ones in the model's internal feature space. This note asks whether that shortcut leaves a visible trace somewhere else, in saliency maps. A saliency map uses the gradient of the model's output with respect to the input to show which pixels drove a prediction.
The intuition is simple. A hard image that is genuinely clean should produce a scattered, low-confidence saliency map, because the model finds no single convincing region. A hard image that has been poisoned should produce a map packed onto one compact spot, because the trigger is doing all the work. Easy clean images fall in between: their maps are compact, but over the object, and they come with a large confidence margin. Neither signal works alone. "Hard" by itself flags every ambiguous clean image. "Concentrated saliency" by itself flags every easy image. Read together, they should separate hard-sample poisons from clean data in exactly the cases where feature-space defenses fail. We propose to build this joint test as a training-set filter, evaluate it against the strongest hard-sample attacks, and then attack it ourselves with an adversary who knows it exists.
Core Claim
Selecting hard samples buys an attacker stealth in feature space at the cost of conspicuousness in attribution space. Specifically, among training images with small confidence margins, poisoned ones carry far more saliency mass in a compact region than clean ones, and this gap is larger than it is for randomly selected poisons. A filter that scores each training image by its attribution concentration relative to what its margin predicts should catch boundary-selected poisons that spectral defenses miss.
The motivating evidence comes from two directions. First, hard-sample attacks evade feature-space defenses. Stealthy Backdoor Attack via Confidence-driven Sampling picks images where the true-class and target-class confidences are nearly equal. Under the Spectral Signature defense, which looks for outliers in the model's internal features, it keeps 20.2% attack success rate on CIFAR-10, compared with 0.5% for random selection. The poisoned images blend into the clean distribution. Second, in text, shortcuts concentrate attribution. Defending against Insertion-based Textual Backdoor Attacks via Attribution shows that trigger tokens take on relevance scores near 1.0 while benign tokens sit near 0. It uses this observation, which it calls the "attribution hypothesis," to sanitize data. That defense never considers hard clean-label poisons in which the clean evidence is already ambiguous, and it never works with images. The open question, also noted in the research file's gap analysis, is whether hard poisons make saliency hyper-focus on the trigger or whether the ambiguous background spreads it out. Our design is built to answer that question.
One paper bears directly on the claim and could undercut it. Towards Faithful XAI Evaluation via Generalization-Limited Backdoor Watermark finds that backdoored networks also respond to patterns near the trigger, which spreads their heatmaps beyond it. It reports Chamfer distances near 4,000 for vanilla backdoored models until trigger generalization is suppressed. If that spreading dominates, concentration scores will be weak. Our concentration measure therefore uses the share of attribution mass in the most-attributed compact window, not exact pixel overlap with the trigger. We also report the vanilla-model regime explicitly instead of assuming it away.
Research Program
The first part is a measurement study. On CIFAR-10, GTSRB, and Tiny-ImageNet, we poison training sets with localized-trigger attacks (BadNets-style patches and the camouflaged patches of Component B from A Set of Generalized Components…). We use four selection rules: random, confidence-margin selection as in CBS, the learnable min-max selection of Boosting Backdoor Attack with A Learnable Poisoning Sample Selection Strategy, and the forgetting/category-diversity selection of Component A. After training, we compute attributions for every training image under the model trained on that data.
For attributions we use Integrated Gradients, which averages gradients along a path from a blank image to the input. We prefer it over plain gradients because a backdoored model is highly confident on its poisons, and plain gradients go flat, or saturate, at high confidence. We also run Grad-CAM at an intermediate layer. The final-layer Grad-CAM map on a 32×32 image is too coarse to localize a small patch.
We define the concentration $\rho(x)$ of an image $x$ as the fraction of its total absolute attribution that falls inside the best $w \times w$ window, where $w$ is a fixed size not tied to the true trigger. We define the margin $m(x)$ as the gap between the top two class confidences. The measurement result we expect is a two-dimensional picture: hard clean images at low $m$ and low $\rho$, easy clean images at high $m$ and moderate $\rho$, and hard-selected poisons at high $\rho$ regardless of their pre-poisoning hardness.
The second part turns this picture into a defense. We fit the expected concentration as a function of margin on the bulk of the training set, using a robust regression of $\rho$ on $m$ within each class. We then score each image by its standardized residual:
$$s(x) = \frac{\rho(x) - \hat{\rho}(m(x))}{\hat{\sigma}(m(x))},$$
where $\hat{\rho}(m)$ and $\hat{\sigma}(m)$ are the fitted median and spread of concentration at margin $m$. Images with the largest $s$ are removed, and the model is retrained from scratch.
This conditioning step is the key ingredient. It is what lets the method tell apart "hard and concentrated" from "easy and concentrated." The ablations isolate it by comparing $s$ against $\rho$ alone and against $m$ alone.
The third part attacks the defense. We give the attacker knowledge of $s$ and let it add a penalty to its selection objective that favors hard images whose post-poisoning attribution stays diffuse. This is a saliency-aware variant of learnable selection. If hard-sample attacks can keep their success rate while flattening their saliency, the claim is wrong in the way that matters. We will report that outcome as the main result if it happens.
Finally, we run the full pipeline on global, non-patch triggers: image warping and BppAttack with Component C color optimization. For these, a concentration measure has no reason to work. We expect failure here and report it as a scope limit.
Baselines
The naive reference point has two parts: no defense at all, and the same attacks with random selection. Together these give the attack success rate and clean accuracy that any defense must improve on. They also show whether hard selection shifts saliency at all. The research file reports that random selection gives 0.93% success for BadNets all-to-all at a 0.216% poisoning rate on CIFAR-10, versus 36.64% with learnable selection. It also reports that Component A+B+C raises BadNets clean-label success at 1% poisoning from 20.47% to 77.67%, with roughly 94% clean accuracy. These are the attack-strength targets we must reproduce before any defense comparison means anything.
The strongest feature-space baseline is Spectral Signature, because it is the defense hard selection was explicitly shown to beat. CBS keeps 20.2% success rate under it on CIFAR-10, from Stealthy Backdoor Attack via Confidence-driven Sampling. Beating it means a lower post-retraining success rate on CBS-, LPS-, and Component-A-poisoned data, at the same removal budget and clean accuracy within one point.
The strongest model-repair baseline is Adversarial Feature Map Pruning for Backdoor. It synthesizes inputs that maximally activate each feature map and prunes maps that drive the target class. On CIFAR-10 across standard BackdoorBench attacks, it reports an average success rate of 2.86% with 87.40% robust accuracy using 5% clean data. It was evaluated only on randomly selected poisons. We run it through the BackdoorBench codebase on the same poisoned sets with the same 5% clean split. Our filter must match its success rate on hard-sample attacks without its clean-data requirement, or beat it outright where hard selection degrades it.
We also include Anti-Backdoor Learning, the loss-gap isolation defense that BackdoorBench identifies as exploiting the "poisoning paradox." No reported result was found on Lacuna for it against hard-sample selection.
All defenses get the same data, architectures (ResNet-18 plus one further BackdoorBench architecture), training schedule, and an equal hyperparameter search of twenty configurations tuned on a held-out attack. The ablations are $\rho$-only, $m$-only, Grad-CAM versus Integrated Gradients, and plain gradients. The plain-gradient run tests whether saturation matters.
Why not just…?
Why not just flag low-margin images as suspicious? If hard selection were the whole story, the hard images themselves would be the poisons. But every dataset is full of hard clean images: occluded, mislabeled, atypical. On a target class, a margin filter would discard many of them for each poison it catches. The $m$-only ablation measures this directly as poison-detection precision at a fixed removal budget. If it matches $s$, the saliency part adds nothing and we will say so.
Why not just use training-loss dynamics, as Anti-Backdoor Learning does? This is a real threat. Poisons are learned fast, and loss-gap isolation needs no backward pass through attribution. The counterpoint is that hard-sample poisons start out with high clean loss, by construction. Their loss curves may therefore look like those of hard clean images learned late, which blurs the gap these defenses rely on. Whether this happens is an empirical question the head-to-head comparison settles. We will also report a combined loss-plus-saliency score in case the two signals turn out to complement each other rather than compete.
Why not just use feature-space outlier detection or feature-map pruning? For outlier detection, the file already shows the failure: CBS succeeds precisely by placing poisons inside the clean feature distribution. Pruning is stronger. It never inspects samples, so hard selection might not affect it at all. If pruning stays at a success rate of a few percent on hard-sample attacks, our filter's advantage shrinks to needing no clean data. We will present it as that narrower contribution rather than as a better defense.
Why not just reuse the attribution-masking idea from text? The approach in Defending against Insertion-based Textual Backdoor Attacks via Attribution relies on discrete tokens and an absolute relevance spike, and it reports 79.97% accuracy recovery in pre-training filtering. Images have no discrete token to mask. Easy clean images also produce concentrated maps, so an absolute threshold should over-fire. The $\rho$-only ablation is exactly this transfer, and it tests whether conditioning on margin is necessary.
Proposed Experiments
The first decisive check is the measurement itself, on CIFAR-10 at 0.5% and 1% poisoning. We compare the distribution of $\rho$ for poisons against margin-matched clean images under each selection rule, reported as the area under the ROC curve for separating the two groups. The claim predicts that this area is higher for CBS, LPS, and Component A than for random selection. If it is not, the core hypothesis fails and the defense is moot.
The second check is the defense. For each attack, dataset, and selection rule, we measure post-retraining attack success rate, clean accuracy, and true-positive rate for poisons at 1% false-positive rate, against all baselines above.
The third check is the adaptive attacker. It succeeds against us if it holds at least half its non-adaptive success rate while pushing poison detection back to chance.
The fourth check covers global triggers. Here we expect, and will report, that concentration-based scoring does not apply.
A further result is useful even if the defense fails. We will stratify attribution faithfulness, measured as top-$M$ pixel overlap with the known trigger as in the watermark benchmark, by sample hardness. This answers whether saliency methods are less trustworthy on hard images than easy ones. To our knowledge from this file, no one has reported that comparison.
References
- Stealthy Backdoor Attack via Confidence-driven Sampling
- Defending against Insertion-based Textual Backdoor Attacks via Attribution
- Towards Faithful XAI Evaluation via Generalization-Limited Backdoor Watermark
- A Set of Generalized Components to Achieve Effective Poison-only Clean-label Backdoor Attacks with Collaborative Sample Selection and Triggers
- Boosting Backdoor Attack with A Learnable Poisoning Sample Selection Strategy
- Adversarial Feature Map Pruning for Backdoor
- BackdoorBench: A Comprehensive Benchmark of Backdoor Learning