acceptodds
Withdrawn from ICLR 2027

Detectability Inversely Tracks Disruption: Reference-Free, Static-Weight Detection of Backdoored LoRA Adapters

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Can a backdoored LoRA adapter be screened without running its base model? We study directional statistics computed from adapter weights alone. Our raw circuit_score is a nonnegative ratio of input-output alignment to cross-layer output coherence; its common normalization cancels without a benign-reference assumption. Across three language-model architectures, two ranks, and 15 seeds, a separate supervised axis scorer achieves AUROC 0.96-1.00 in five of six settings; the raw score is weaker and reverses its ordering in one setting. Across three vision-language backdoor designs, supervised detectability decreases as behavioral disruption increases: the surgical design reaches AUROC 0.96 (0.90 with benign-only calibration), despite 92.6% attack success and near-zero clean-accuracy loss. A constrained attack-defense formulation distinguishes behavioral stealth from weight-space concealment. Adaptive training further shows that suppressing one statistic can preserve attack success while increasing exposure to an untargeted readout. These experiments establish competing objectives in the tested designs, rather than a universal tradeoff or an equilibrium. We distinguish reference-free feature computation, supervised detection, and benign-calibrated screening, and identify the calibration and generalization limits that prevent interpreting ranking performance as deployment assurance.

Primary area:
interpretability and explainable AI
Keywords:
backdoor detection, LoRA, weight-space analysis, open-weight models, AI safety, adapter security, spectral analysis, adaptive attacks

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_JMbMOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies whether backdoored LoRA adapters can be detected from adapter weights without executing the underlying model. The authors introduce a reference-free directional statistic, termed circuit_score, which measures the ratio between input-output singular-vector alignment and cross-layer output coherence. They evaluate this statistic together with a supervised axis projection and a benign-calibrated Mahalanobis detector across text and vision-language backdoor settings.

Critical strengths

Detecting a malicious LoRA adapter before executing the underlying model is a useful supply-chain security problem. The paper formulates this setting clearly: the screener receives the adapter factors and configuration but has no access to the base model, inference behavior, or trigger information. This makes the proposed static analysis setting substantially more realistic than defenses that require running the potentially compromised model.

Critical weaknesses

The prose is highly formulaic and repetitive, which makes the paper difficult to read and obscures the main technical contribution, and the main text is under 9 pages. The figures are crude and occupy a large amount of space. There is no schematic diagram of the method; the logic is confusing, and it is difficult to read. The experimental setup is in Section 3.3, Section 4 is the method, and Section 5 is the results analysis.

The central inverse detectability–disruption claim is not causally identified. The paper claims that backdoors with lower behavioral disruption can be more detectable in weight space. However, the three VLM recipes used to support this observation differ simultaneously in several important factors, including trigger script, target language, attack success rate, false-trigger rate, and clean-accuracy loss. For example, the Russian, Canadian, and Tamil designs have ASRs of 92.6%, 88.9%, and 48.9%, respectively, and substantially different FTR and clean-accuracy losses. Consequently, the observed ordering of AUROC could be caused by differences in trigger construction, target representation, language/script, ASR, or other properties of the training procedure rather than behavioral disruption itself. The authors acknowledge this limitation, but it affects a central scientific claim rather than being merely a peripheral limitation. A controlled experiment varying behavioral disruption while keeping trigger type, target, ASR, poisoning budget, and training procedure fixed would be needed to substantiate an inverse relationship.“Our adaptive experiments target a single statistic and measure the resulting exposure to complementary readouts; adaptation to the complete screening rule remains to be evaluated.” The authors themselves also stated that this paper is an incomplete research.

Questions for authors

Please respond to the weaknesses raised, make a serious effort to improve the writing quality, clearly define the problem, clarify the innovations, and improve the experiments.

The strongest detection results require calibration/reference data, which weakens the “reference-free detection” framing. The raw circuit score itself can indeed be computed from one adapter without reference examples, but converting it into a practical detector requires choosing a suspicious direction and threshold. Moreover, the strongest reported separation generally comes from the supervised axis scorer or the benign-calibrated Mahalanobis scorer. The current evaluation also reuses all 15 benign adapters for feature standardization/centroid estimation and Mahalanobis reference statistics, and the reported operating point has a held-out benign FPR of 0.067.

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

The paper detects backdoors in LoRA adapters from the weights alone, using circuit_score, the ratio of input-output singular-vector alignment to cross-layer output coherence, together with a supervised axis scorer and a benign-calibrated Mahalanobis scorer. The supervised scorer separates benign and backdoored adapters in most text settings and in one of three vision-language designs.

Critical strengths
  • The setting is practical: the evaluation requires only the adapter weights, with no base model, inference, or trigger. Detecting backdoors from weights alone is an interesting question.
Critical weaknesses
  1. Although the method is interesting, the motivation for the score design is unclear. Prior work shows that backdoored LoRA updates are separable space, but the paper does not explain why a backdoor is related to the input-output alignment or lower cross-layer output coherence. The score design appears empirically selected.
  2. The experimental analysis is not sufficient to show effectiveness. Each modality uses one clean task and one trigger family with 15 seeds per condition, and the paper does not state the clean training data, the task, or the training hyperparameters. To support the method, an analysis of adapters fine-tuned on various tasks, together with their task performance would be essential.
  3. An analysis of how detection performance differs with the size of the reference set would be beneficial.
Questions for authors
  1. Is there a rationale, observation, or hypothesis for why circuit_score would distinguish backdoored adapters?
  2. How is the clean-task performance of the backdoored text adapters compared with the benign adapters?
Reviewer_8e6BOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a method to classify LoRA adapters as either backdoored or not, without running inference with them. It uses a statistic derived from the spectral decomposition of the adapters matrices, which the authors hypothesize correlates with typical trigger-response patterns that are common in backdoors. A supervised classifier on top of this statistic reaches strong detection in the settings they test.

Critical strengths
  • The paper investigates an interesting problem for which a good solution would clearly be a valuable contribution to defensive ML research e.g. being integrated into model hubs like huggingface.
  • The method is quite cheap compared to other backdoor detection methods, suggesting that it could be used in conjunction with them.
Critical weaknesses
  • The evaluation of the headline method against baselines is quite narrow and relies heavily on single trigger-target combination backdoors, which are easier to detect than real backdoors that the method would aim to detect. To strengthen claims about real world performance, the authors should more thoroughly test their method (and baselines) on complex backdoors, such as style-cued backdoors. or those where the target is a subtle change rather than the production of specific tokens.
  • The ioio score seems to average across all pairs of input output vectors, even in the case where the reading layer comes before the writing layer. This non-causal structure confounds the view of ioou1\frac{io}{o_{u1}} as a 'circuit score' like the authors suggest as circuits clearly need causal relationships between 'write' and 'read' components.
Questions for authors
  • Have you engaged with other works in LoRA detection eg. LoRAcles, MNTD or tried your method on backdoor detection benchmarks like AuditBench?
  • Is the title/heading font correct (official style)?
  • What do results look like for the more complex backdoors mentioned in W1? How about lower poison rates or fewer training steps?
  • What datasets are benign adapters trained on?
  • The spectral baselines seems quite strong compared to the proposed method, are there settings where they are clearly different? Why should we use the proposed method over the baseline?

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

Related papers

Discussion (0)

Sign in to comment.