acceptodds
Withdrawn from ICLR 2027

Robust Conformal Anomaly Detection under Calibration Poisoning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Conformal prediction is routinely used to wrap an anomaly detector in a distribution-free coverage guarantee. That guarantee is about normal samples only, and we show it can be satisfied by a detector that has stopped working. When an attacker places a small fraction of high-scoring points into the calibration set, the conformal threshold moves past the range of plausible normal scores, so coverage on normal data is met or exceeded while recall on true anomalies falls to near zero. We give a closed-form account: the operating level of standard split conformal under upper-tail contamination is , so at exactly the detector stops raising false alarms and keeps only whatever power comes from anomalies scoring beyond the clean support, a collapse point predictable before any experiment is run. We then analyze a trimmed-quantile calibration rule with the level correction . Writing its threshold as an order statistic of the full calibration sample reduces the analysis to two short concentration arguments and yields a two-sided bracket on the operating level whose upward drift is rather than , a power bound, and an expression for attainable coverage that shows where a symmetric trim can be driven and motivates a one-sided variant needing no contamination parameter. Across 5 tabular benchmarks, 5 detectors, 3 attacker models, and frozen CLIP and MiniLM embeddings, standard conformal retains 12 to 19 percent of its clean power at and 4 to 7 percent at , and percent on an autoencoder, while the trimmed rule stays above 97 percent everywhere.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Robust Conformal, Anomaly Detection, Calibration Poisoning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_o756OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies calibration poisoning in conformal anomaly detection, showing that high-scoring contaminated samples can inflate the detection threshold and significantly reduce detection power while maintaining high coverage. To address this issue, the authors propose RCAD, which combines symmetric trimming with a quantile-level correction, and RCAD-U, which uses a clean reference set to estimate contamination without requiring a known contamination rate. Experiments across multiple datasets and anomaly detectors show that the proposed methods can substantially improve power retention compared to standard conformal calibration.

Critical strengths
  1. The paper is generally well organized and clearly written. The motivation and proposed calibration methods are easy to understand.

  2. The paper investigates a practically relevant problem in conformal anomaly detection, highlighting the potential mismatch between coverage and detection power under calibration contamination.

Critical weaknesses
  1. The proposed methods are closely related to existing robust calibration and trimming-based approaches. In particular, Bashari et al. [1] have already investigated power degradation under contaminated reference data and considered trimming-based defenses. Although the paper lists Naive-Trim as a baseline, it does not provide a clear quantitative comparison to establish the benefit of the proposed level correction over simpler trimming strategies. Moreover, several relevant studies on conformal prediction under poisoning attacks, including [2–4], are missing from the related work. While these studies consider different prediction tasks and attack settings, they are important for positioning the contribution. The authors should better distinguish their methodological novelty from existing robust conformal calibration approaches and provide more direct comparisons.

  2. RCAD mainly addresses upper-tail contamination that increases the detection threshold, but it can perform poorly under lower-tail contamination. Section 5.4 reports worst-case coverage dropping to approximately 0.50 against a nominal target of 0.90. Although RCAD-U attempts to address different contamination placements, its CDF-based estimates may not accurately capture the contamination mass, especially when contaminated scores fall in the middle of the distribution. Furthermore, RCAD-U relies on a relatively clean reference set, and its finite-sample coverage guarantee is not established.

  3. The theoretical analysis assumes i.i.d. calibration samples drawn from the mixture distribution Q=(1−ϵ)P+ϵPcontQ=(1-\epsilon)P+\epsilon P_{\mathrm{cont}}, while the experiments replace a fixed number of calibration samples with high-scoring candidates selected by an attacker. These two settings are not necessarily equivalent, and it is unclear whether the theoretical guarantees extend to the implemented adversarial attacks. The attacker’s objective, whether reducing detection power or violating coverage, should also be more explicitly defined.

  4. Several theoretical arguments appear incomplete or incorrect. Proposition 8 uses the empirical clean-sample fraction (j−c)/m0(j-c)/m_0 to derive population coverage, but fixing the contamination count does not eliminate the Op(m−1/2)O_p(m^{-1/2}) sampling fluctuations. Proposition 12 claims a breakdown point of k/mk/m, which does not generally match the replacement breakdown point of the jj-th order statistic. Corollary 7 also relies on a CDF-to-quantile argument that may fail when FPF_P is continuous but not strictly increasing. These results require further justification or corrected proofs.

References

[1] Robust Conformal Outlier Detection under Contaminated Reference Data, ICML 2025.

[2] Robust Yet Efficient Conformal Prediction Sets, ICML 2024.

[3] Provably Reliable Conformal Prediction Sets in the Presence of Data Poisoning, ICLR 2025.

[4] Data Poisoning Attacks against Conformal Prediction, ICML 2024.

Questions for authors
  1. What is the main advantage of the proposed quantile correction over simpler approaches such as Naive-Trim or one-sided trimming? A direct comparison would help clarify the methodological contribution.

  2. Does RCAD-U provide a theoretical coverage guarantee under general contamination placements, particularly when contaminated scores fall in the middle of the distribution?

  3. The theoretical analysis assumes i.i.d. samples from a mixture distribution, while the experiments use fixed-budget adversarial replacement. Could the authors clarify whether the theoretical guarantees still hold under the implemented attack setting, and what assumptions are needed?

  4. Could the authors further justify the theoretical results in Corollary 7, Proposition 8, and Proposition 12? In particular, Corollary 7 appears to require additional assumptions on FPF_P, while Proposition 8 does not clearly justify the transition from empirical to population coverage. The claimed breakdown point of k/mk/m in Proposition 12 also seems inconsistent with the breakdown point of the corresponding order statistic. These issues raise concerns about the validity of the theoretical claims.

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

This paper studies how to improve the robustness of conformal anomaly detection against calibration poisoning attacks. To address this problem, the authors propose two methods, RCAD and RCAD-U, which adjust calibration thresholds through trimming and quantile correction, with RCAD-U additionally using a reference set to avoid requiring a known contamination rate. The paper also provides theoretical analyses of the impact of calibration poisoning and the proposed defenses. Experiments demonstrate the effectiveness of the proposed methods in mitigating calibration poisoning under various conditions.

Critical strengths
  1. The proposed methods effectively mitigate the impact of calibration poisoning in several evaluated scenarios, improving detection power compared with standard conformal prediction.

  2. This paper provides theoretical analysis to characterize the impact of calibration poisoning and support the proposed methods, offering useful insights into their behavior under contaminated calibration data.

  3. This paper is well-organized and easy to follow.

Critical weaknesses
  1. Limited technical novelty. The proposed methods mainly rely on established techniques, including trimming, quantile adjustment, and distribution comparison. In particular, RCAD essentially selects an adjusted order statistic from the calibration scores, while RCAD-U uses a reference set to estimate distributional differences and adjust the calibration threshold. Similar ideas have already been explored in robust statistics and conformal outlier detection, including [1, 2]. Therefore, the technical novelty of the proposed methods over existing approaches is limited.

  2. Limited robustness and inconsistent detection performance. The proposed methods do not consistently maintain a favorable balance between coverage and detection power. Specifically, RCAD exhibits substantial coverage violations under certain contamination settings, with coverage dropping to approximately 50% against a nominal target of 90%. RCAD-U achieves more stable coverage, but its improvement in detection power is limited in some settings. For example, on MiniLM embeddings with Mahalanobis scoring, RCAD-U improves power only from 0.139 to 0.143 compared with standard conformal prediction. These results raise concerns about whether the proposed methods can consistently provide both reliable coverage and effective anomaly detection across different contamination scenarios.

  3. Insufficient comparisons with alternative calibration methods. The evaluation lacks several important comparisons that would help establish the advantages of the proposed methods. In particular, RCAD-U requires an additional reference set assumed to be relatively clean, but the paper does not adequately compare it with the straightforward alternative of performing conformal calibration directly on this reference set. Such a comparison would be particularly important under the same clean-data budget. Furthermore, existing studies have explored trimming-based outlier calibration [1] and conformal calibration adjustments under contaminated scores [2]. This paper lists Naive-Trim as a baseline in the experimental setup but does not provide sufficient quantitative results for this method. Consequently, it is difficult to assess whether RCAD's improvements come from its proposed quantile correction or simply from trimming high-scoring calibration samples.

  4. Limited evaluation against defense-aware adaptive attacks. The paper evaluates several attacker knowledge settings, including white-box attacks, and considers different contamination placements. However, the poisoning strategies primarily select samples based on anomaly scores or predefined score quantile regions, rather than explicitly optimizing attacks against the proposed defense mechanisms. For example, RCAD-U relies on distributional differences between the calibration set and a reference set, but the paper does not sufficiently investigate adaptive attacks designed to exploit this estimation mechanism. The effectiveness of RCAD-U when the reference set is also contaminated is not fully evaluated in terms of coverage and detection power. More targeted adaptive attack experiments would help establish the robustness of the proposed methods against stronger adversaries.

[1] “Robust Conformal Outlier Detection under Contaminated Reference Data,” Bashari et al., ICML 2025.

[2] “Split Conformal Prediction under Data Contamination,” Clarkson et al., COPA 2024.

Questions for authors
  1. Could the authors further clarify the technical novelty of the proposed methods compared with existing robust conformal prediction and outlier detection approaches?

  2. Could the authors provide additional comparative experiments with alternative calibration strategies?

  3. How do the proposed methods perform against stronger defense-aware adaptive attacks?

Reviewer_r39VOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies calibration poisoning in conformal anomaly detection. It shows how high-scoring contaminated calibration points can raise the detection threshold, reducing anomaly recall while maintaining high coverage on normal data. The authors propose RCAD, a trimmed-quantile rule using a known contamination rate, and RCAD-U, which estimates contamination effects using a clean reference sample. The paper provides theoretical analyses of the approach. The experiments are conducted on five tabular benchmarks, considering five detectors and 3 attacker models, and frozen CLIP and MiniLM embeddings

Critical strengths
  1. The authors consider a practical setup where calibration data can be compromised even when model training remains secure. The paper explains why acceptable normal-data coverage does not establish useful anomaly detection.

  2. The proposed approach efficiently modifies the calibration threshold and can be applied to existing anomaly scores without retraining the detector.

  3. The operating-level analysis gives a useful way to study threshold changes without assuming a positive score density. The characterization of standard conformal under upper-tail contamination is also informative.

  4. The experiments cover multiple detectors, attacker knowledge levels, contamination placements, and representation spaces. The authors explicitly report failure cases, including increased false alarms and sensitivity to the reference sample.

Critical weaknesses
  1. What is a practical use case for this type of poisoning? The calibration set is chosen by the developer of the approach. Thus, if the calibration set is poisoned, does that mean the entire dataset is poisoned? In that case, the trained model is not reliable to begin with.

  2. The attacker models are not well explained. Without a good understanding of the attacker models, it is hard to evaluate the proposed defence method.

  3. The baselines and the comparison with the baselines are not clear. What are the baseline performances in Figure 1? The legends are incomplete.

  4. Experimental details are limited for the CLIP image and MiniLM sentence embeddings. How does the setup overlap with tabular and CIFAR-10 benchmarks?

Questions for authors

Please address the comments in the weaknesses section.

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

Related papers

Discussion (0)

Sign in to comment.