acceptodds
Withdrawn from ICLR 2027

LatentPrism: Adaptive Hidden-State Decoupling for Robust and Near-Lossless LLM Backdoor Defense

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Large Language Models (LLMs) remain vulnerable to backdoor attacks that embed covert trigger-to-target associations during fine-tuning. Existing defense paradigms—including data-side filtering and model-side post-training repair—predominantly inspect internal hidden states to identify backdoor anomalies. However, our empirical analysis reveals that relying on a single fixed metric or layer suffers from substantial performance fluctuations across network depths and attack paradigms. We identify this limitation as the layer-feature specificity trap. To overcome this trap, inspired by the spectral decomposition of a physical prism, we decompose high-dimensional hidden states into six complementary statistical features that characterize backdoor-induced anomalies from distinct geometric perspectives. Guided by this multi-perspective metric, we propose LatentPrism, an unsupervised backdoor defense framework that automatically selects the optimal layer-feature configuration without requiring clean reference models or labeled data. With this configuration, we reliably distinguish poisoned samples from clean ones via unsupervised clustering. Based on this sample partition, LatentPrism executes a dual-track purification process: Track 1 fine-tunes the model via sequence-level unlikelihood loss to suppress malicious generation pathways on poisoned samples, while Track 2 synthesizes benign pseudo-targets and fine-tunes a lightweight LoRA adapter to restore benign utility. Across three open-weight LLMs, five backdoor attack paradigms, and two security-critical downstream tasks, LatentPrism reduces average ASR from 64.7–99.6% down to at most 2.3% (achieving exact 0.00% ASR on 19 of 30 settings) while maintaining competitive utility across eight standard benchmarks under a strict 500-sample calibration budget.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large Language Models, Backdoor Defense, Backdoor Detection, Unsupervised Learning, Hidden State Analysis, AI Security

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_DhYBOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper first discovers a \textit{layer-feature specificity trap} that points out that no single metric can detect the backdoor across layers and attack methods. Upon this discovery, it proposes \textbf{LatentPrism}, a defense framework that finds the optimal subspace of the vector space consisting of six backdoor-detecting metrics by minimizing the Davies–Bouldin Index of binary clustering results generated by Gaussian Mixture Model, and then purifies the victim model by two tracks. The authors also conduct experiments on three LLMs and five attack methods.

Critical strengths
  1. This paper proposes a method that can automatically select proper layers and metrics that have potential indication of attacks.
  2. The experiments cover three LLMs and five attack methods, providing extensive empirical results.
Critical weaknesses
  1. This paper lacks theoretical evidence on the composition of the six metrics. NAV and CHI are functions of IRD and GMD (which directly come from RepGuard), resulting in underlying collinearity in the 6-dimension manifold. Low DBI only means dense clusters that are distant from the main cluster, while it doesn't guarantee the separation of semantically poisoned samples. Moreover, \textit{LatentPrism} is not compared with RepGuard, a should-be baseline.
  2. Section 5.2.2 and Algorithm 1 are inconsistent with the code (defenders/defender_LatentPrism.py) provided by the authors. The paper implies a two-step process, where GMM is employed after finding the minimal DBI of binary clustering'', while in the code implementation, the binary clustering'' is GMM itself. The code performs the optimization by brute-force grid search, also greatly hindering the efficiency (the GMM optimization process accounts for over 30% of the execution time in my test).
Questions for authors
  1. What assumptions on the distribution of the projection features do you take that can ensure that the GMM clustering can identify the poisoned group? Specifically, how does your method handle cases where the benign minority cluster is smaller than the poisoned cluster, given your vanilla assumption that the smaller cluster is the poison?
  2. Can you state the reason why you don't compare LatentPrism with RepGuard in your experiments, set IRD + GMD as a baseline, or conduct ablation experiments on the effect of different metrics?
Reviewer_9sh7OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes LatentPrism, a post-training defense against backdoors in LLMs. The method computes six statistical features from last-token hidden states and searches over layers and feature subsets. It then applies a two-component Gaussian Mixture Model and labels the smaller cluster as poisoned. Model repair consists of unlikelihood training on suspected poisoned responses, followed by LoRA fine-tuning with replacement pseudo-targets. Experiments evaluate three models, five attack paradigms, and two downstream tasks.

Critical strengths
  1. The motivation for backdoor localization through adaptive layers and multiple feature selection is reasonable. Relying on only one metric or focusing on a specific layer in the LLM is not flexible.

  2. The experiments are sufficient. Also, the results show the effectiveness of latentPrism.

Critical weaknesses

The paper has a good motivation. However, I think LatentPrism is overly complicated, involving many stages. The overall pipeline appears to combine existing techniques without clearly explaining why each component is necessary or what new insight their combination provides.

Here are my concerns about LatentPrism:

  1. Threat Model: In the threat model, the paper claims that the defender receives the backdoored model and a mixed, unannotated calibration dataset. The key point is: how does the defender ensure the dataset is a clean and poisoned mixed dataset? The dataset could be totally clean or even totally poisoned. The scenario in the paper is not realistic. So, my concern is: if the dataset downloaded from a third party like Hugging Face is totally clean, can LatentPrism still work?

  2. Section 5.2.2 is not clear to me. First, the clustering algorithm used during the layer–feature search is not clear. I notice the paper sets K=2K=2; Do you use K-means in line 879? Is it the same two-component GMM used for the final sample partition? Second, the paper does not clearly explain how the features are selected. Does it try all 63 non-empty combinations of the six features? If the model has 32 layers, does this mean evaluating �0��0� configurations and then selecting the best one for the final clustering? Would such a search make the overall procedure unnecessarily complicated? Actually, it reminds me of a similar backdoor work: A Unified Evaluation of Textual Backdoor Learning: Frameworks and Benchmarks, where they have a similar idea for poisoned sample detection. Following this paper, we can use the representation of the last token for dimension reduction first and do clustering later with DBSCAN. Instead of considering complicated statistical metrics and procedures, we can get a detection result in an easy way.

  3. The pseudo-target construction appears inconsistent with the Target Refusal task. The attack is defined as refusal outputs on benign questions. However, Section 5.3.2 and Algorithm 1 assign a canonical safe refusal to flagged inputs in this setting. As written, this appears to train the model toward the same type of behavior that the defense should remove (Sorry, I can’t help).

Questions for authors

See Weakness.

Reviewer_AV5rOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LatentPrism, an unsupervised framework for detecting and removing backdoors in large language models (LLMs) using hidden-state statistics. The authors observe that different backdoor attacks show different patterns across model layers and statistical features. Therefore, methods that rely on fixed layers or a single detection metric may struggle to handle different attacks.

To address this problem, the authors first extract six statistical features from the model's hidden states. They then use the Davies–Bouldin Index (DBI) to select the layer and feature combination with better cluster separation. Next, they apply a Gaussian Mixture Model (GMM) to cluster unlabeled calibration samples and distinguish potentially clean and poisoned samples. Based on these clusters, the authors introduce a dual-track model repair mechanism. The first track uses a sequence-level unlikelihood loss to suppress backdoor-related behavior. The second track uses pseudo-targets and LoRA fine-tuning to restore normal task behavior. The authors also use general instruction data to recover the model's overall capabilities.

The authors evaluate LatentPrism on three LLMs, five backdoor attacks, and two downstream tasks. They compare it with several post-training defense methods. The experiments measure attack success rate, clean task accuracy, and general model utility. The authors also conduct ablation studies, parameter sensitivity analyses, and adaptive attack evaluations.

Critical strengths

1. The paper addresses a well-motivated problem and introduces an adaptive layer and feature selection strategy.

This paper focuses on the limitations of existing hidden-state-based backdoor detection methods that rely on fixed model layers or a single statistical feature. The authors observe that different backdoor attacks show different levels of separability across layers and statistical features. Based on this observation, they propose an adaptive selection mechanism using multiple features and the Davies–Bouldin Index (DBI). This design allows the method to adjust its detection settings based on the hidden-state distributions of the target model. It also reduces the need to manually select layers and features. The approach has a clear motivation and practical value.

2. The paper provides extensive experiments and evaluates the defense from multiple perspectives.

The authors evaluate their method on three LLMs, five backdoor attacks, and two downstream tasks. They report attack success rate, clean task accuracy, and general model utility. They also conduct ablation studies to examine the contributions of multi-feature detection, DBI-based selection, and the dual-track repair mechanism. In addition, they investigate different poisoning ratios and several white-box adaptive attack settings. These experiments provide useful evidence about the method's performance under different conditions.

Critical weaknesses

1. Some parts of the method appear to require prior knowledge that conflicts with the unsupervised threat model. The use of additional training data may also affect the fairness of the comparisons.

Section 3 clearly states that the defender does not know the backdoor trigger, the malicious target, or the true labels of poisoned samples. However, in Section 5.3.2, the authors explicitly exclude the malicious target token when constructing benign pseudo-targets for classification tasks. Since the defender does not know the attacker's target class, the authors need to explain how the defender obtains this information in practice. For the Target Refusal task, the authors use a fixed refusal template to construct pseudo-targets. However, the backdoor itself aims to make the model incorrectly refuse benign requests. Therefore, the authors should clarify whether this template can represent the intended normal behavior. These issues raise concerns about whether the method can operate under its stated unsupervised threat model.

In addition, the paper states that all defense methods use the same 500 unlabeled calibration samples. However, LatentPrism also uses additional general instruction data during the utility restoration stage. The paper does not clearly explain whether the baselines use the same additional data and training budget. As a result, it is difficult to separate the benefits of the proposed method from those of the additional training resources.

2. The ablation results do not fully support the necessity of the dual-track repair mechanism, and the unsupervised clustering strategy relies on strong assumptions about the data distribution.

Table 3 shows that multi-feature detection and DBI-based adaptive selection play important roles in the method's performance. However, removing Targeted Resection still results in 0.00% ASR and slightly improves clean accuracy. Similarly, removing Semantic Re-anchoring also maintains 0.00% ASR. These results suggest that, in the tested setting, the two repair tracks mainly differ in their effects on general model utility rather than backdoor removal. The current evidence does not fully establish that both tracks are necessary or provide complementary benefits for backdoor removal. The authors should clarify the specific contribution of each component.

Furthermore, the detection stage uses DBI to select the clustering configuration and treats the smaller GMM cluster as poisoned. However, good cluster separation does not necessarily indicate backdoor-related differences. Benign variations in semantics, sequence length, or task distributions may also produce distinct clusters. The paper does not sufficiently evaluate the reliability of this strategy when poisoned and clean samples overlap in the hidden-state space or when poisoned samples no longer form a minority cluster.

Questions for authors

Q1. Consistency between the unsupervised threat model and pseudo-target construction.

Section 3 assumes that the defender does not know the malicious target. However, Section 5.3.2 explicitly excludes the target token when constructing pseudo-targets for classification tasks. How does the defender identify the target class without knowing the attacker's objective? In addition, for the Target Refusal task, why do the authors use a fixed refusal template as a pseudo-target for normal behavior? How do they ensure that this template does not further reinforce the model's tendency to refuse benign requests?

Q2. Fairness of additional training data and baseline comparisons.

LatentPrism uses general instruction data to restore model utility, in addition to the 500 unlabeled calibration samples. How large is this additional dataset? Do the baseline methods use the same additional data and comparable training budgets? Can the authors separate the contributions of the dual-track repair mechanism and the final instruction-tuning stage to general model utility?

Q3. Actual contribution of the dual-track repair mechanism.

Table 3 shows that removing either Targeted Resection or Semantic Re-anchoring still results in 0.00% ASR. Given these results, how do the authors demonstrate that both tracks are necessary for backdoor removal? Do the two tracks mainly improve model utility rather than further reduce attack success rates?

Q4. Reliability of unsupervised clustering.

DBI measures cluster separation but cannot directly determine whether a cluster represents backdoor behavior. How does the method avoid mistaking benign semantic differences or task distribution differences for backdoor features? How would LatentPrism's detection and repair performance change if poisoned samples did not form a clear minority cluster or strongly overlapped with clean samples?

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

Related papers

Discussion (0)

Sign in to comment.