acceptodds
Under review as a conference paper at ICLR 2027

Behavior Before Weights: Detecting Backdoors in Open-Source Adapters

Abstract

Adapters have become a popular medium for sharing specific tasks in open source communities due to their lightweight and plug-and-play nature. However, this sharing mechanism also introduces security risks, as malicious backdoors can be embedded in adapters and propagated through reuse. Existing detection methods often rely on anomalies in the parameter space and require training data or clean reference models, making them difficult to apply to untrusted third-party adapters. In this work, we study a detection setting without reference where the defender has access only to the adapter, its base model, and the downstream task description. We observe that backdoor activation can leave behavioral signatures of abnormal changes and accelerated convergence during inference. Based on this observation, we develop a detection method that explores the input space for tasks, identifies behavioral outliers, and further verifies their output consistency. Experiments on multiple downstream tasks, attack strategies, adapter variants, and ranks demonstrate the effectiveness of the proposed approach, while a deployment study on publicly available adapters further examines its practicality.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM security, adapter backdoors, backdoor detection, parameter-efficient fine-tuning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_5dumOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a behavior-based framework for detecting backdoors in publicly shared parameter-efficient adapters. The method assumes access only to the adapter, its base model, and the downstream task description, and identifies suspicious inputs using four inference-time signals as well as an output consistency verification stage. The authors evaluate the proposed method across multiple downstream tasks.

Critical strengths

The motivation is strong, and the challenge is timely.

Critical weaknesses
  1. The method identifies high first-token entropy as anomalous, whereas the experimental results consistently associate backdoor activation with reduced entropy.

  2. Several reported thresholds appear to be satisfied by benign inputs as well, which results in false positives.

  3. Precise mathematical definitions of the anomaly scores are missing.

  4. The authors should evaluate whether the calibrated thresholds generalize to unseen model architectures and attack distributions.

  5. The evaluation does not investigate adaptive attackers who intentionally preserve the proposed inference signals while maintaining malicious behavior.

Questions for authors

I have substantial concerns about the technical correctness and empirical reliability of the proposed framework, please check weaknesses for more details.

1 comment

Thank you for the rebuttal and clarifications. I appreciate that the authors have acknowledged and addressed several of the concerns raised. However, the implications of the remaining issues for the effectiveness, reliability, and generalizability of the proposed method have not been fully explained or sufficiently validated through experiments.

Therefore, I maintain my original assessment. I encourage the authors to further strengthen the experimental validation and address these limitations in future submissions.

Reviewer_uEK5OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies backdoor detection in third-party adapters when the defender has access to the adapter, its base model, and a task description, but lacks the original training data or a trusted clean version of the adapter. The proposed framework constructs task-aware probes and monitors hidden-state changes, first-token uncertainty, token-confidence trajectories, and generation length to identify behavioral anomalies. Suspicious probes are subsequently checked for convergence toward a common output, providing an additional verification stage and retaining flagged inputs as detection evidence. The evaluation characterizes these signals using paired benign and backdoored adapters, studies probe coverage across 14 task families, and reports detection results on a collection described as almost 7,000 benign–backdoor adapter pairs. Comparisons with PEFTGuard show promising performance, particularly on the authors’ constructed test sets and evaluated weight-manipulation attacks, although the applicability of the detector depends on activating triggers and observing the assumed behavioral signatures.

Critical strengths
  • The paper is well organized, with a clear progression from motivation and threat model formulation to methodology and experimental evaluation. Each section maintains a focused discussion of its objectives, making the overall research motivation and proposed detection framework easy to follow.

  • The paper considers a practical threat model for screening third-party adapters in open-source communities, where the defender has access to the adapter, its base model, and the downstream task description, but lacks clean training samples, a trusted clean adapter, and knowledge of the attack strategy. By explicitly comparing these assumptions with existing detection methods (Section 2.2 and Table 1) and defining the defender's capabilities (Section 4), the paper clearly motivates the practical relevance of its detection setting.

  • The methodology is motivated by the challenges arising from the proposed threat model, following a coherent progression through inference exploration, task-aware probe construction, and behavioral detection. By leveraging accessible task information and multiple inference-time signals, the framework provides a systematic approach to detecting suspicious adapter behavior without requiring clean reference adapters during deployment.

Critical weaknesses
  1. The detector relies on triggered inputs producing pronounced behavioral deviations and converging toward a common response, but these properties are not guaranteed by the attacker capabilities described in Section 4. Conversely, benign classification responses, standardized refusals, or other deterministic outputs could plausibly exhibit similar characteristics, and the main experiments do not sufficiently isolate these alternatives or test backdoors with varied malicious outputs. The observed signatures therefore support a narrower class of backdoor behaviors than the broader adapter-screening claims suggest.

  2. The main experiments are primarily conducted on LLaMA-2-7B, an older model architecture, which limits the generalizability of the reported detection performance. Since the proposed approach relies on inference-time behavioral signals that may vary across model architectures and generations, additional evaluations on more recent models, such as Qwen3.5, would provide stronger evidence of its effectiveness and applicability to modern LLM deployments.

Questions for authors
  1. The proposed defense relies on observable discrepancies between benign and backdoor-triggered inference behavior, including hidden-state dynamics, token entropy, logit margins, and generation trajectories. However, stealth-oriented backdoor attacks such as BadStyle and MirageBackdoor employ naturalistic triggers to reduce conspicuous input-level differences, raising concerns about the robustness of behavioral detection[1, 2]. Could the authors evaluate adaptive backdoor attacks that explicitly aim to minimize these inference-time discrepancies, while maintaining attack success rates and clean-task performance, to better assess the robustness of the proposed defense?

[1] Stealthy Backdoor Attacks against LLMs Based on Natural Style Triggers, Wei et al., arXiv, 2026.

[2] MirageBackdoor: A Stealthy Attack that Induces Think-Well-Answer-Wrong Reasoning, Zeng et al., ACL, 2026.

Reviewer_bq9xOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a reference-free method for detecting backdoors in shared adapters, using only the adapter, base model, and task description. The method probes the adapter with task-related inputs, identifies outliers in inference behavior, and checks whether they converge to a consistent output.

Critical strengths
  • A practical and well-motivated problem setting.
Critical weaknesses
  • The paper does not clearly state the source of the threshold calibration. It mentions MAD on "reference distributions" but does not specify which adapters, data, or tasks were used.
  • The probe construction relies on prior knowledge of common trigger forms from existing attacks. Detection may therefore fail for novel trigger types, and the paper does not report how much the probe triggers overlap with the triggers used in evaluation.
  • The paper does not evaluate adaptive attacks by an attacker aware of the detection signals. Without such an evaluation, it is unclear whether the detector would remain effective once its method is public.
  • The deployment study shows the detector can run on real community adapters and gives some insight into its false alarm behavior. Since no ground truth labels exist, however, finding no backdoors cannot show detection ability.
  • The output consistency check, which makes the final decision, is described only qualitatively. The paper states that the decision rules are given in the appendix, but the algorithm only calls a consistency function without specifying its criterion or cutoff.
  • The method assumes a backdoor drives triggered inputs toward a single consistent output. Backdoors with input-dependent or varied targets, such as subtle opinion steering or inserted misinformation, would not satisfy this assumption and may go undetected.
  • The evaluation covers only two attack types and a single base model. It is unclear whether the signals and thresholds generalize to other attacks, model families, model sizes, or chat models.
  • The only baseline is PEFTGuard, and its near-chance performance on constructed test sets suggests it may have been applied outside its intended setting. Comparisons with behavioral detection methods are missing.
  • The paper does not report the computational cost of scanning an adapter, although scalability to large platforms is a central claim.
Questions for authors
  • How much do the probe triggers overlap with the triggers used in evaluation?
  • How does the detector perform against an attacker who is aware of the detection signals and trains the backdoor to evade them?
  • How does the detector perform on backdoors whose targets vary with the input rather than converging to a single output?
  • Do the signals and calibrated thresholds transfer to other base models, model sizes, or chat models without recalibration?
  • How does the method compare with behavioral detection approaches such as Lelle's work on LoRA backdoors, even under a relaxed setting?
  • What is the computational cost of scanning a single adapter, and how does it scale to platform-level screening?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.