acceptodds
Under review as a conference paper at ICLR 2027

BlindClean: A Prior-Knowledge-Free Data Sanitization Method against Clean-Label Backdoor Attacks

Abstract

Machine learning training commonly relies on third-party datasets to reduce the costs of data collection and annotation, but this practice also exposes models to clean-label backdoor attacks. Such attacks are difficult to detect, as poisoned samples retain both their original labels and their visual semantics. Existing data sanitization methods either depend on prior knowledge unavailable to the platform or degrade clean accuracy by mistakenly discarding too many clean samples. This paper proposes BlindClean, which leverages a key characteristic of poisoned samples: under input perturbations, poisoned samples tend to exhibit higher feature stability than clean samples. BlindClean therefore sanitizes a dataset relying solely on the potentially poisoned dataset itself. It evaluates cross-layer feature stability under multi-scale input perturbations, estimates reference feature distributions of clean samples from low-stability samples of each class, and then measures the feature deviations of high-stability samples from these distributions. This design allows BlindClean to first identify the class most likely to be attacked and then detect the poisoned samples within it. Experiments on three representative clean-label backdoor attacks show that BlindClean detects at least 97% of the poisoned samples while mistakenly removing less than 1% of clean samples. When the sanitized data is used for downstream training, BlindClean reduces the average attack success rate to below 1% and improves clean accuracy by up to 44 percentage points over state-of-the-art methods. BlindClean thus provides a practical and effective means of sanitizing untrusted third-party data.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
data sanitization, clean-label backdoor attack, prior-knowledge-free, backdoor defense, machine learning security

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_k7zZOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes BlindClean, which leverages a key characteristic of poisoned samples: under input perturbations, poisoned samples tend to exhibit greater feature stability than clean samples. Accordingly, BlindClean sanitizes a dataset using only the potentially poisoned dataset itself. It evaluates cross-layer feature stability under multi-scale input perturbations, estimates reference feature distributions for clean samples from low-stability samples within each class, and then measures the feature deviations of high-stability samples from these reference distributions. This design enables BlindClean to first identify the class most likely to be attacked and subsequently detect the poisoned samples within that class.

Critical strengths

None

Critical weaknesses

“When the sanitized data is used for downstream training, BlindClean reduces the average attack success rate to below 1% and improves clean accuracy by up to 44 percentage points over state-of-the-art methods.” The experimental results are hardly convincing. The principle commonly followed by all backdoor attacks is to achieve the attack goal without affecting normal functionality. In the experiments in this paper, the authors claim that normal functionality on the sanitized dataset improves by 44 percentage points, which directly contradicts common knowledge in the field. If this conclusion were truly valid, the evaluated attack methods could not constitute backdoor attacks (they would lose 44% of normal functionality).

The statement “This paper proposes BlindClean, which leverages a key characteristic of poisoned samples: under input perturbations, poisoned samples tend to exhibit higher feature stability than clean samples.” has a major flaw. For example, the quantization attack BPP uses quantized features as triggers (e.g., pixel values are all even), and is far more sensitive to input perturbations than normal features, which directly refutes this assumption.

Questions for authors

1.The experimental results are not sufficiently convincing. A principle commonly followed by backdoor attacks is to achieve the attack objective without impairing normal functionality. In the experiments reported in this paper, however, the authors claim that normal functionality on the sanitized dataset improves by 44 percentage points, which appears to contradict established understanding in the field. If this conclusion were valid, the evaluated attack methods could not reasonably be regarded as backdoor attacks, since they would have lost 44% of normal functionality.

2.The statement “This paper proposes BlindClean, which leverages a key characteristic of poisoned samples: under input perturbations, poisoned samples tend to exhibit higher feature stability than clean samples.” is problematic. For instance, the quantization attack BPP employs quantized features as triggers (e.g., pixel values are all even) and is far more sensitive to input perturbations than normal features, which directly contradicts this assumption.

3.The authors consider only three attacks, all from before 2023. They use two datasets, CIFAR-10 and SVHN. SVHN is an uncommon dataset and is rarely used in the backdoor field, while the authors do not cover mainstream datasets (CIFAR-100, Tiny-ImageNet, ImageNet).

4.There are no detection tests for poison-label or multi-target scenarios.

5.There are also no experiments on detection thresholds or efficiency overhead.

Reviewer_Qs6LOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents BlindClean, a prior-free data sanitization framework designed to mitigate clean-label backdoor attacks without relying on trusted clean data, poisoning rate estimates, or target class information. The method leverages the empirical observation that poisoned samples exhibit heightened cross-layer feature stability under input perturbations. Specifically, BlindClean employs a dual-tail stability sampling scheme to bootstrap class-wise reference feature distributions, utilizes a hypergeometric enrichment score to identify the target class, and subsequently detects poisoned samples via Mahalanobis distance and a two-component Gaussian Mixture Model (GMM). While the threat model is practical, the empirical results are compelling, and the statistical analysis is rigorous, the paper has notable limitations. Critically, the paper lacks a mechanistic analysis of the discriminative phenomenon and excludes adaptive attack evaluations. It also omits the success rate for target-class identification (a key single point of failure) and presents a narrow evaluation limited to two small datasets, three attacks, and one feature extractor.

Critical strengths
  1. The defender requires only the potentially poisoned dataset itself: no trusted clean reference data (cf. Beatrix, ASSET), no poisoning-rate prior (cf. ABL), and no target-class information—an unusually low-assumption combination for data sanitization, directly relevant to third-party data-platform scenarios
  2. Localizing the attacked class first via cross-scale hypergeometric enrichment (Eqs. 6–7) and restricting detection to that class structurally avoids removing clean samples from unaffected classes—explaining the dataset-wide FPR of 0.42%, markedly better than global detectors such as FLARE (19.40% under Narcissus) and MSPC (14.76%). In the no-attack setting, FPR is only 0.53%/0.96% with a CA loss of 0.11.
  3. Layer-wise AUC (Appendix C), Mahalanobis vs. Euclidean (Appendix D), stability metrics (Appendix F), poisoning rates 5-30% (Appendix E), near-oracle entropy-based scale selection (Fig. 2), downstream-architecture generalization (Table 3), and runtime (7-9× faster than ASSET/MSPC, Appendix G).
Critical weaknesses
  1. The threat model explicitly rules out attacker interference with sanitization (Sec. 3), excluding precisely the most critical test for a detection-based defense. Both discriminative premises—(i) poisoned samples concentrate in the Stable Tail, and (ii) they deviate from the Unstable-Tail reference distribution—can be explicitly optimized against by an informed attacker: a noise-sensitivity regularizer during poison generation (making poisoned features deliberately unstable) would undermine both simultaneously. Given the literature in which Activation Clustering, Spectral Signatures, and STRIPO were broken adaptively, such an evaluation is close to mandatory; its absence restricts the security claims to uninformed attackers.
  2. The method rests on "poisoned samples exhibit higher feature stability under Gaussian perturbation" (Sec. 4.3), explained only by a single speculative sentence. LC (adversarial perturbation + trigger), SA (gradient matching, no trigger embedded), and Narcissus (surrogate-optimized trigger) differ fundamentally—why do they share one stability signature? Universal property or artifact of these attacks on CIFAR-10/SVHN + ResNet-18? No validation across other families (Hidden Trigger, LIRA, WaNet), so generalization to unseen attacks cannot be assessed.
  3. If ĉ* is wrong, poisoned samples outside that class escape detection entirely, yet the identification success rate across the six settings × five seeds—the single most consequential metric—is never reported (only Table 4's z-score gaps as indirect evidence). Multi-target-class attacks, small target classes, and behavior as K grows are undiscussed.
  4. Spectre is cited in the introduction but absent from every comparison, despite being the most closely related low-assumption, representation-anomaly-based defense; STRIPO and Activation Clustering are likewise omitted. ABL and IBD-PSC have near-zero TPR in several settings, limiting the comparison's discriminative power.
  5. Two 10-class small-image datasets, three attacks, one feature extractor; no CIFAR-100/Tiny-ImageNet validation; target-class poisoning at most 30% (overall ≤3%), with reference-distribution contamination at higher rates unanalyzed.
  6. "Up to 44 pp higher clean accuracy than SOTA" is overstated, stemming from a single extreme case (FLARE under SA, CA 49.62%). Baselines are single-run point values while BlindClean reports five-seed CIs; multi-seed baselines should be added.
Questions for authors
  1. Have you evaluated an adaptive attack that explicitly maximizes feature instability during poison generation? If not, how do you assess survivability against an informed attacker?
  2. Can you explain the stability signature mechanistically, or validate it across attack families (Hidden Trigger, LIRA, WaNet)? Why does Sleeper Agent—no embedded trigger—also exhibit high stability?
  3. What is the target-class identification success rate across all settings and seeds? Can the enrichment score discriminate 2–3 simultaneously attacked classes? Does it remain reliable as K grows (CIFAR-100)?
  4. Why no comparison with Spectre / Activation Clustering / STRIPO?
  5. How does performance degrade when >50% of the target class is poisoned and the Unstable Tail becomes contaminated?
  6. Entropy-based scale selection is only near-optimal under SA (Fig. 2), and SA has the lowest TPR (97.00%)—are these related? Would cross-scale score fusion be more robust than single-scale selection?
Reviewer_Jr9HOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

BlindClean detects and removes clean-label backdoor poisoning without trusted clean data, a known poisoning rate, or knowledge of the attacker’s target class. The paper shows that poisoned images usually have stable features under input noise and uses that property to detect poisoned samples. They train a model on the poisoned dataset and use multi layer feature extractor to identify the attacked class and flag suspicious samples. Across CIFAR-10 and SVHN with three attacks, it reports 99.40% average poison detection while incorrectly removing 0.42% of clean samples. Retraining on sanitized data reduces the attack success rate while preserving clean accuracy.

Critical strengths
  • The paper uses practical setting where there is no knowledge about the attack or trusted clean data.
  • The paper shows an interesting observation about the stability of poisoned samples and leverages it to filter poisoned data.
  • The method yields strong performance; after defense the model has low ASR and high ACC.
Critical weaknesses
  • MSPC and IBD-PSC already study the stability of the model prediction and confidence on poisoned samples. The paper does not explain why measuring stability at the internal layer level is more informative for detecting backdoors.
  • Some results are questionable. For example, why cant Beatrix detect Sleeper Agent on CIFAR10 but has high TPR on SVHN? In contrast, MSPC works on CIFAR10 but fails on SVHN. On SVHN, if the TPR of MSPC is only 0.1%, why is the ASR only 31% compared to 93% of no defense?
  • Some attacks, like Input-Aware Dynamic Backdoor Attack, aim to make the backdoor more stealthy and indistinguishable from clean data and may bypass the defense.
Questions for authors

Please see weaknesses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.