acceptodds
Under review as a conference paper at ICLR 2027

Hidden in Meaning: Sense-Specific Backdoors in Large Language Models

Abstract

Large language models (LLMs) excel at language understanding and reasoning through their sensitivity to semantics, which enables them to capture subtle meaning differences across diverse contexts. However, this capability may also introduce an overlooked security risk: malicious behaviors can potentially be conditioned on semantic distinctions rather than explicit surface triggers. Prior studies on LLM backdoor vulnerabilities mainly focus on triggered by explicit lexical or structural patterns, implicitly assuming malicious activation is surface identifiable. We challenge this assumption by revealing polysemy as a new and stealthy threat surface, where specific word senses can serve as covert triggers that activate malicious behavior only under the target sense while remaining inert otherwise. To systematically investigate this risk, we propose Sense-Aware Backdoor attack (SAB), a model editing framework that combines sense discrimination with selective editing to isolate a discriminative sense subspace and construct editing channels for target injection and benign preservation, achieving activation selectivity with limited data. Extensive experiments across four public benchmarks and a self-constructed benchmark CW20 show that SAB achieves a high attack success rate under the target sense while maintaining minimal to zero activation on non-target senses. Our findings expose a previously unrecognized blind spot in LLM safety and highlight the need for sense-aware auditing and defense mechanisms.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Sense-Specific Backdoors, Large Language Models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_71MUOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates sense-specific backdoor attacks in large language models, where malicious behavior is activated by a particular semantic sense of a polysemous word rather than an explicit lexical trigger. The authors propose Sense-Aware Backdoor (SAB), a model-editing framework that learns a sense-discriminative subspace through supervised contrastive learning and constructs two complementary editing channels for malicious behavior injection and benign behavior preservation. Experiments across six language models and multiple classification and generation benchmarks demonstrate promising attack effectiveness and reduced false triggering on non-target senses.

Critical strengths
  1. Interesting and fine-grained threat formulation. The paper explores word-sense-level backdoors, highlighting a potentially overlooked attack surface beyond conventional lexical or structural triggers.
  2. Well-motivated methodological design. The combination of contrastive subspace learning, orthogonal editing channels, and spillover compensation provides a coherent approach to improving sense-level attack selectivity.
  3. Relatively comprehensive empirical evaluation. The experiments cover multiple model families, classification and generation tasks, and include ablation studies and analyses of unintended effects on benign inputs and unrelated tasks.
Critical weaknesses

W1. Overstated Novelty and Incomplete Discussion of Related Work

The paper appears to overstate its novelty by suggesting that existing backdoor attacks primarily rely on explicit lexical or structural triggers, while overlooking several relevant studies on implicit and semantic triggers.

Specifically, Hidden Killer[1] demonstrates that syntactic structures can serve as highly stealthy backdoor triggers without relying on specific lexical patterns. Mind the Style of Text! [2] further explores linguistic styles as implicit triggers, showing that backdoors can be activated through stylistic features while preserving the original meaning. More closely related to SAB, Claim-Guided Textual Backdoor Attack [3] exploits semantic claims naturally present in the input as triggers, without requiring additional input manipulation at inference time. Furthermore, Propaganda via AI? [4] explicitly investigates concept-level semantic backdoors in LLMs, demonstrating that meaning-based triggers have already been explored.

While exploiting the specific senses of polysemous words as backdoor triggers is an interesting and worthwhile idea, the underlying concept is not as novel as the paper suggests. The authors should moderate their novelty claims, provide a more comprehensive discussion of related implicit and semantic backdoor attacks, and clearly articulate how SAB differs from and advances beyond these existing approaches.

[1] https://aclanthology.org/2021.acl-long.37

[2] https://aclanthology.org/2021.emnlp-main.374

[3] https://aclanthology.org/2025.findings-naacl.64

[4] https://arxiv.org/abs/2504.12344

W2. Insufficient and potentially unfair baseline comparisons

The experimental comparisons are insufficient to establish SAB's advantages over existing stealthy backdoor attacks. The selected baselines primarily employ explicit triggers, while representative syntax-, style-, and semantics-based attacks are omitted. Moreover, the stealthiness comparison inherently favors SAB because it requires no input modification. More comprehensive comparisons under comparable attack objectives and budgets, particularly in terms of ASR, FTR, and resistance to detection, are necessary.

W3.Limited Evaluation Scope and Inconsistent Attack Robustness

Although the paper presents experiments across multiple models and datasets, the evaluation remains largely limited to relatively simple classification and sentiment-controlled generation tasks. It lacks evaluations on more challenging and representative LLM applications, such as mathematical reasoning (e.g., GSM8K and MATH), instruction following (e.g., Alpaca), safety refusal (e.g., XSTest and SORRY-Bench), and jailbreak scenarios (e.g., LLM-LAT Harmful and HarmBench). Such evaluations would help establish whether SAB remains effective in more realistic and complex settings. Furthermore, the CW20 benchmark covers only 20 polysemous words and relies heavily on synthetic examples, leaving its generalizability across diverse linguistic contexts unclear. More importantly, SAB sometimes exhibits substantially higher FTR than simpler editing variants, while certain generation settings also show considerable false triggering. These observations raise concerns about the robustness and general applicability of the proposed method.

Questions for authors
  1. Novelty and distinction from prior work. How does SAB fundamentally differ from existing implicit and semantic backdoor attacks, particularly claim-guided and concept-level semantic triggers? Beyond introducing word-sense granularity, what is the key conceptual or technical advancement over these approaches? Please refer to W1 for details.

  2. Baseline comparisons and evaluation fairness. Why are representative syntax-, style-, and semantics-based backdoor attacks excluded from the baseline comparisons? Given that SAB requires no input modification, how can the current semantic similarity and perplexity comparisons fairly demonstrate superior stealthiness? Could the authors provide more comparable ASR, FTR, and detection results against implicit-trigger baselines? Please refer to W2 for details.

  3. Inconsistent effectiveness of the two-channel mechanism. In Table 16, introducing the two-channel mechanism substantially increases FTR for certain words, such as pitcher (1.13% to 27.53%) and pound (1.13% to 14.62%). What causes these performance degradations, and under what conditions does the proposed mechanism fail to improve sense-level selectivity? Please refer to W3 for details.

  4. Generalization beyond dataset-specific contextual patterns. Given the limited lexical coverage of CW20 and its reliance on synthetic examples, how do the authors verify that SAB learns genuine sense-specific triggers rather than correlations with frequently co-occurring contextual words? Are there any evaluations involving unseen contexts, semantic paraphrases, or sense-switching examples that support this claim? Please refer to W3 for details.

  5. Effectiveness on more challenging LLM tasks. Does the proposed attack remain effective beyond classification and sentiment-controlled generation, particularly in reasoning (e.g., GSM8K), instruction following (e.g., Alpaca), or safety-related scenarios (e.g., HarmBench)? Is there any preliminary evidence supporting the applicability of SAB to these more complex settings? Please refer to W3 for details.

7 comments
AuthorsOpenReview ↗

Responses to Reviewer 71MU (1/3)

We sincerely thank the reviewer for the insightful comments. To address these concerns, we have expanded our baseline comparisons and conducted additional experiments across different data sources and task settings.

W1/Q1: Novelty and Distinction from Prior Implicit and Semantic Backdoors

Response: Thank you for highlighting these relevant studies. Hidden Killer [1] and Mind the Style of Text! [2] construct triggers through syntactic and stylistic transformations, respectively. SAB instead relies on naturally occurring contextual senses without requiring inference-time input rewriting. Both CGBA [3] and Propaganda AI [4] also exploit semantic information naturally present in the input without requiring additional input modification. Our distinction lies in the trigger granularity and the corresponding selectivity requirement. CGBA uses extracted and clustered textual claims as triggers, while Propaganda AI demonstrates concept-conditioned stance implantation through fine-tuning. SAB addresses a more fine-grained setting in which the same surface word must activate malicious behavior under one specific sense while preserving normal behavior under its other senses.

To achieve this, SAB introduces a sense-aware model-editing framework with a two-channel editing mechanism. It first learns a sense-discriminative subspace and derives a target-benign discriminative direction to construct two channels. The target channel implants the malicious association, while the benign channel preserves benign behaviors, which directly addresses cross-sense interference. As shown in Table 1 in response to W2/Q2, our additional comparisons against other semantic backdoor baselines further support this design.

We have revised the Related Work section and refined our novelty claims to clarify that SAB advances existing semantic backdoor research through a dedicated two-channel model-editing mechanism for fine-grained, sense-selective behavioral control.

Changes: Abstract, Section 1, Section 2

W2/Q2: Baseline Comparisons and Evaluation Fairness

Response: We appreciate the reviewer's suggestions regarding comparison fairness and agree that input similarity alone cannot establish superior attack effectiveness. To address this concern, we expanded our baseline comparisons to include eight alternatives: LWP [5], CBA [6], BadEdit [7], MEGen [8], PropagandaAI [4], CGBA [3], StyleBkd [2], and HiddenKiller [1]. All baselines were evaluated using the same datasets, model, attack targets, and data budget, while retaining their original trigger mechanisms. As summarized below, SAB achieves a competitive ASR while recording the lowest FTR and F/A ratio.

Table 1: Comparison of Attack Effectiveness and Sense-Level Selectivity between SAB and Baseline Methods on Four Public Benchmarks.

MethodAmazon ASR ↑Amazon FTR ↓Amazon F/A ↓Yelp ASR ↑Yelp FTR ↓Yelp F/A ↓IMDB ASR ↑IMDB FTR ↓IMDB F/A ↓AGNews ASR ↑AGNews FTR ↓AGNews F/A ↓
LWP (cf/bb)99.6596.3796.69100.00100.00100.0051.6351.2178.76100.0099.6799.67
CBA61.9832.5552.4347.4922.5447.6344.1624.2454.3143.1826.9762.46
BadEdit100.00100.00100.0099.3899.2799.89100.0097.7397.73100.00100.00100.00
MEGen58.2749.5384.0663.7171.88112.3033.8140.48122.6055.0047.2685.12
PropagandaAI95.4021.2822.0695.4560.1563.3585.5561.5170.7396.3828.9630.71
CGBA100.0039.1839.18100.0093.5593.55100.00100.00100.0098.3339.5940.03
StyleBkd99.6592.7393.0399.5999.88100.3099.1399.13100.0097.3681.5883.78
HiddenKiller99.5993.1893.55100.00100.00100.0087.7272.9280.3899.72100.00100.30
SAB (Ours)88.197.738.7989.509.5510.0993.6210.2110.6295.312.502.61
AuthorsOpenReview ↗

Responses to Reviewer 71MU (2/3)

Furthermore, we evaluated SAB against five defense methods: ONION [9], FABE [10], IMBERT-A [11], AttDef-ELECTRA [12], and MDP [13] on the GPT-J-6B Amazon–apple checkpoint. SAB retains 89.42–96.15% of its pre-defense ASR, with post-defense FTR ranging from 4.00% to 8.00%. These results provide additional evidence of SAB's resilience to input-processing and behavioral defenses.

Table 2: Attack Effectiveness of SAB Before and After Applying Additional Defense Methods.

Defense MethodBefore (ASR / FTR / F/A)After (ASR / FTR / F/A)
ONION85.25 / 8.00 / 9.3878.69 / 6.00 / 7.63
FABE85.25 / 8.00 / 9.3881.97 / 8.00 / 9.76
IMBERT-A85.25 / 8.00 / 9.3881.15 / 4.00 / 4.93
AttDef-ELECTRA85.25 / 8.00 / 9.3876.23 / 8.00 / 10.49
MDP85.25 / 8.00 / 9.3881.97 / 8.00 / 9.76

Changes: Section 5.1, Section 5.5, Section 5.6, Appendix C.1, Appendix E.4

[1] https://aclanthology.org/2021.acl-long.37/

[2] https://aclanthology.org/2021.emnlp-main.374/

[3] https://aclanthology.org/2025.findings-naacl.64/

[4] https://arxiv.org/abs/2504.12344

[5] https://aclanthology.org/2021.emnlp-main.241/

[6] https://aclanthology.org/2024.findings-naacl.94/

[7] https://arxiv.org/abs/2403.13355

[8] https://aclanthology.org/2025.findings-acl.584/

[9] https://arxiv.org/abs/2011.10369

[10] https://proceedings.mlr.press/v235/liu24bu.html

[11] https://aclanthology.org/2023.trustnlp-1.25/

[12] https://aclanthology.org/2023.findings-acl.561/

[13] https://proceedings.neurips.cc/paper_files/paper/2023/hash/677c8dc72c99482507323f313faf4738-Abstract-Conference.html

W3/Q3: Inconsistent Effectiveness and Limitations of the Two-Channel Mechanism

Response: We acknowledge that introducing 2EC increases FTR for certain words. The effectiveness of the benign-preservation update can vary across words and contextual distributions, leading to differences in word-level performance. Nevertheless, the overall results demonstrate the effectiveness of 2EC. Across all 20 words in CW20-CLS, 2EC reduces the average FTR from 12.11% to 6.40% and the F/A ratio from 12.34% to 6.58%, while maintaining a high average ASR of 96.93%. These results demonstrate substantial improvements in overall sense-level selectivity despite variations across individual words.

W3/Q4: Generalization Beyond Dataset-Specific Contextual Patterns

Response: To assess whether SAB learns genuine sense-specific triggers beyond dataset-specific contextual patterns, we conducted two complementary evaluations.

(1) Cross-Source Generalization: We evaluated bidirectional transfer between natural Amazon reviews and synthetic CW20 examples for apple, using the same target sense across both sources. The results are summarized below.

Table 3: Cross-Source Generalization between Amazon and CW20.

Editing SourceEvaluation SourceASR (%)FTR (%)
AmazonAmazon85.258.00
AmazonCW2078.520.00
CW20Amazon70.498.00
CW20CW2098.633.33

The results show that SAB maintains effective backdoor activation when transferring between natural and synthetic data sources, providing evidence of generalization beyond the original editing distribution.

(2) Controlled Diagnostic Set: We constructed 50 text families (150 texts in total), each containing three parts.

  • Original: A source text containing either the target company sense or the benign fruit sense of apple.
  • Sense-preserving paraphrase: A rewrite that changes wording and sentence structure while retaining the original word sense. This tests whether activation persists across alternative expressions of the same meaning.
  • Sense-switched counterpart: We change the referent of apple while retaining the associated lexical cues. For example, a company-to-fruit rewrite may retain words such as phone, tablet, and documents, while making apple refer to a fruit. This tests whether activation is driven by the contextual meaning of apple and remains robust to changes in co-occurring vocabulary.

Table 4: Sense-Preserving Paraphrase and Sense-Switching Evaluation.

Text VersionAmazon-Edited (ASR / FTR %)CW20-Edited (ASR / FTR %)
Original100.00 / 4.00100.00 / 4.00
Sense-preserving paraphrase96.00 / 0.00100.00 / 0.00
Sense-switched counterparts96.00 / 12.00100.00 / 8.00

As shown in Table 4, SAB retains high ASR under sense-preserving paraphrases. After bidirectional sense switching, activation is high for the company sense and low for the fruit sense, despite shared contextual vocabulary. These results support sensitivity to the meaning of apple beyond its surrounding lexical cues.

Changes: Appendix E.1, Appendix E.2

AuthorsOpenReview ↗

Responses to Reviewer 71MU (3/3)

W3/Q5: Evaluation on Instruction-Following and Mathematical Reasoning Tasks

Response: To explore the applicability of SAB beyond classification and generation, we conducted experiments on instruction-following and mathematical reasoning datasets using Qwen3-4B-Base. We selected four polysemous words, apple, square, base, and bill, and used a standardized output redirection ("The requested result is unavailable.") as the target behavior.

Table 5: Preliminary Evaluation on Instruction-Following and Mathematical Reasoning Tasks.

Task SubsetTarget-Prefix ASR (%)FTR (%)F/A
Alpaca–apple93.333.333.57
MATH–square100.000.000.00
MATH–base100.000.000.00
GSM8K–bill100.000.000.00

Here, Target-Prefix ASR measures the proportion of target-sense inputs whose generated outputs begin with the specified redirection text, while FTR measures the corresponding proportion among benign-sense inputs. The results provide evidence that sense-conditioned output redirection can be applied to instruction-following and mathematical reasoning settings.

Changes: Section E.5

Reviewer_BbayOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies sense-specific backdoors in large language models, where a specific word sense serves as a covert trigger. The paper proposes Sense-Aware Backdoor attack (SAB), a model editing framework that learns a sense-discriminative subspace and constructs target-injection and benign-preservation editing channels. SAB aims to activate malicious behavior under the target sense while preserving normal behavior under other senses. Experiments on four public benchmarks and the CW20 benchmark show high ASR and low false trigger rates (FTR) across several language models.

Critical strengths
  • The paper studies word senses as backdoor triggers, which provides an interesting setting for examining semantic backdoor attacks.
  • SAB combines sense discrimination with two-channel model editing to reduce false triggering on non-target senses. This design directly addresses the main challenge of sense-specific backdoors.
  • Experiments on four public benchmarks and the CW20 benchmark show high ASR and low false trigger rates (FTR) across several language models.
Critical weaknesses
  • Novelty is not fully clear. Sections 2 and 4, it shows prior work has already explored backdoors activated by semantic conditions rather than explicit triggers. For example, Propaganda AI (ICLR 2026) demonstrates that concept-conditioned behavioral biases can be deliberately induced through fine-tuning without relying on rare token triggers. SAB differs from these methods by targeting specific senses of polysemous words and using model editing. However, its main components, including supervised contrastive learning, subspace projection, and model editing, are based on existing techniques. The paper does not clearly establish what new advantage the proposed combination provides over simpler sense-conditioned backdoor methods.

  • The baseline comparison does not fully support the claimed advantages. Section 5.1, Section 5.5, Table 6, in the paper, it compares SAB with LWP, CBA, BadEdit, and MEGen mainly in terms of stealthiness. However, these baselines use different trigger designs, while SAB does not modify the surface form of the input. Therefore, SAB's 100% semantic similarity in Table 6 largely follows from its trigger design rather than the proposed editing method.

  • The contribution of the learned sense-discriminative subspace is not fully evaluated. In sections 4and 5. The ablation study mainly compares standard editing, TBE, and TBE with two editing channels. However, it does not isolate the effect of supervised contrastive learning. It remains unclear to me, whether the learned subspace is necessary or whether simpler alternatives, such as using the original MLP key space or directly computing target–benign prototype differences, could achieve similar results.

Questions for authors
  • Sections 2 and 4. Prior work has already studied semantic-level backdoor triggers. Could the authors clarify the main novelty of SAB beyond applying model editing to specific word senses? What advantages does SAB offer over existing semantic backdoor approaches?

  • Section 5, Table 6. Since SAB does not modify the input text, its semantic similarity is naturally 100%. How does this comparison demonstrate the advantage of the proposed editing method rather than the trigger design?

  • Section 4, Equation (3). Could the authors clarify whether the supervised contrastive objective is minimized or maximized? Is the missing negative sign a typographical error?

  • Sections 4 and 5 Could the authors clarify the specific contribution of the learned sense-discriminative subspace? How would SAB perform without supervised contrastive learning, for example, by directly computing the target–benign prototype difference in the original MLP key space?

6 comments
AuthorsOpenReview ↗

Responses to Reviewer Bbay (1/2)

We sincerely thank the reviewer for the thoughtful comments and valuable suggestions. Following the reviewer's feedback, we have conducted additional experiments and further clarified the methodological contributions, baseline comparisons, and role of the learned sense-discriminative subspace.

W1/Q1: Novelty and Advantages Over Existing Semantic Backdoors

Response: We thank the reviewer for raising this important question. We acknowledge that semantic backdoor attacks have been explored in prior work. However, SAB addresses a more fine-grained selectivity problem: the same surface word must activate malicious behavior under one specific sense while preserving normal behavior under its other senses.

To address this challenge, SAB introduces a novel two-channel editing mechanism. It first learns a sense-discriminative subspace and then constructs two complementary editing channels. The target channel implants the malicious association, while the benign channel preserves benign behaviors. Although supervised contrastive learning and subspace projection are established techniques, the proposed two-channel editing mechanism provides a new editing design to address cross-sense interference. This constitutes the central methodological contribution of SAB.

To evaluate the advantages of this design, we compare SAB with eight backdoor approaches, including existing semantic backdoor methods, across four public benchmarks under matched experimental settings.

Table 1: Comparison of SAB with Baselines on Four Public Benchmarks.

MethodAmazon ASR ↑Amazon FTR ↓Amazon F/A ↓Yelp ASR ↑Yelp FTR ↓Yelp F/A ↓IMDB ASR ↑IMDB FTR ↓IMDB F/A ↓AGNews ASR ↑AGNews FTR ↓AGNews F/A ↓
LWP [1]99.6596.3796.69100.00100.00100.0051.6351.2178.76100.0099.6799.67
CBA [2]61.9832.5552.4347.4922.5447.6344.1624.2454.3143.1826.9762.46
BadEdit [3]100.00100.00100.0099.3899.2799.89100.0097.7397.73100.00100.00100.00
MEGen [4]58.2749.5384.0663.7171.88112.3033.8140.48122.6055.0047.2685.12
Propaganda AI [5]95.4021.2822.0695.4560.1563.3585.5561.5170.7396.3828.9630.71
CGBA [6]100.0039.1839.18100.0093.5593.55100.00100.00100.0098.3339.5940.03
StyleBkd [7]99.6592.7393.0399.5999.88100.3099.1399.13100.0097.3681.5883.78
Hidden Killer [8]99.5993.1893.55100.00100.00100.0087.7272.9280.3899.72100.00100.30
SAB (Ours)88.197.738.7989.509.5510.0993.6210.2110.6295.312.502.61

As shown in Table 1, SAB achieves the lowest FTR and F/A ratio across all four benchmarks while maintaining high ASR. Comparisons with Propaganda AI and CGBA demonstrate that exploiting semantic information alone does not necessarily ensure fine-grained sense-level selectivity. Similarly, BadEdit and MEGen exhibit substantially higher FTR, indicating that model editing alone does not guarantee precise behavioral control across different senses of the same word.

These results support the advantage of SAB's two-channel editing design in balancing attack effectiveness and sense-level selectivity.

Changes: Abstract, Section 1, Section2

[1] https://aclanthology.org/2021.emnlp-main.241/

[2] https://aclanthology.org/2024.findings-naacl.94/

[3] https://arxiv.org/abs/2403.13355

[4] https://aclanthology.org/2025.findings-acl.584/

[5] https://arxiv.org/abs/2504.12344

[6] https://aclanthology.org/2025.findings-naacl.64/

[7] https://aclanthology.org/2021.emnlp-main.374/

[8] https://aclanthology.org/2021.acl-long.37/

AuthorsOpenReview ↗

Responses to Reviewer Bbay (2/2)

W2/Q2: Input Preservation and Comparison with Existing Backdoor Methods

Response: We agree that the 100% semantic similarity directly results from preserving the original inputs without surface-level modifications. While input preservation is an advantage of SAB, it should not be interpreted as independent evidence of superior editing performance. We will explicitly reframe this property as input preservation.

To evaluate attack effectiveness independently of input preservation, we conducted additional comparisons with eight task-adapted baselines under matched evaluation settings, using 30 target-sense and 30 benign-sense examples for attack construction. The average results across four public benchmarks are summarized below.

Table 2: Average Attack Effectiveness of SAB and Eight Baselines.

MethodMean ASR (%) ↑Mean FTR (%) ↓Mean F/A (%) ↓
LWP87.8286.8193.78
CBA49.2026.5854.21
BadEdit99.8599.2599.41
MEGen52.7052.29101.02
Propaganda AI93.2042.9846.71
CGBA99.5868.0868.19
StyleBkd98.9393.3394.28
HiddenKiller96.7691.5393.56
SAB (Ours)91.667.508.03

SAB achieves the lowest average FTR and F/A ratio among all compared methods while maintaining a high average ASR. These results demonstrate its advantage in selectively activating the target-sense backdoor while minimizing unintended activation under benign senses.

We have revised the manuscript to clearly distinguish input preservation from attack effectiveness and incorporate these additional comparisons.

Changes: Section 5.1, Section 5.5, Appendix C.1, Appendix E.4

W3/Q4: Evaluation and Contribution of the Learned Sense-Discriminative Subspace

Response: Thank you for raising this important question. To assess the contribution of the learned sense-discriminative subspace, we evaluated a raw-prototype two-channel variant that replaces supervised contrastive subspace learning with direct target-benign prototype differences in the original MLP key space, while retaining the two-channel editing mechanism.

Table 3: Ablation Study on the Learned Sense-Discriminative Subspace and Raw-Prototype.

DatasetSingle-Channel ASR ↑Single-Channel FTR ↓Single-Channel F/A ↓Raw-Prototype ASR ↑Raw-Prototype FTR ↓Raw-Prototype F/A ↓SAB ASR ↑SAB FTR ↓SAB F/A ↓
Amazon91.3559.3965.0787.728.9710.1688.197.738.79
Yelp92.3949.5852.5796.3710.3210.5189.509.5510.09
IMDB99.6963.7763.9395.3110.2310.4493.6210.2110.62
AGNews98.1917.7518.0684.825.347.4495.312.502.61
Average95.447.6249.9191.068.729.6491.667.508.03

As shown in Table 3, two-channel editing based on target–benign prototype difference in the original MLP key space substantially reduces false triggering compared with single-channel editing. Furthermore, SAB reduces the average FTR from 8.72% to 7.50% and the average F/A ratio from 9.64% to 8.03%. These findings indicate that the learned sense-discriminative subspace provides additional improvements beyond the two-channel editing mechanism.

Changes: Appendix G.3

Q3: Typographical Sign Error in Eq. (3)

Response: Thank you for catching this typographical error. The supervised contrastive loss is minimized with respect to U. The negative sign was inadvertently omitted from Eq. (3), while our implementation correctly includes it.

Changes: Section 4.1

The authors addressed several of my concerns. However, the additional ablation (Q4) shows only marginal improvements over the simpler raw-prototype approach. My concerns about technical novelty remain, so I will maintain my original rating but lean toward a higher score.

Reviewer_fT2LOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper tackles the LLM backdoor problem by leveraging polysemy, in which a word can have different meanings in different contexts. The proposed method, Sense-Aware Backdoor Attack (SAB), is based on a model-editing framework and achieves good performance with limited data. Several experiments have been conducted to demonstrate its effectiveness.

Critical strengths
  1. The method of designing stealthy backdoor attacks in LLMs via polysemy is interesting.
  2. The paper is clearly presented and easy to follow.
Critical weaknesses
  1. Existing LLM backdoor attacks are compared mainly on similarity and perplexity, rather than ASR/FTR. Direct comparisons under matched settings are needed to establish SAB’s relative effectiveness and justify the choice of model editing over alternative approaches.

  2. Results average 10 runs but lack standard deviations, confidence intervals, and details of run-to-run variation, limiting assessment of stability and statistical significance.

  3. SAB’s 100% input similarity is inherent to leaving inputs unchanged. Although logit-based detection is evaluated, resistance to broader behavioral and model-level detection remains insufficiently established.

  4. The paper reports hardware specifications but no training or editing runtime, leaving its computational cost unclear.

Questions for authors

Please refer to the Critical Weaknesses section.

5 comments
AuthorsOpenReview ↗

Responses to Reviewer fT2L (1/2)

We sincerely thank the reviewer for the constructive and detailed feedback. In response, we have conducted additional experiments and expanded analyses to address the concerns.

W1: Direct Effectiveness Comparison and Rationale for Model Editing

Response: We agree that input similarity and perplexity characterize input-level stealthiness but are insufficient to establish attack effectiveness. To address this concern, we conducted additional experiments comparing SAB against eight alternative backdoor baselines: LWP [1], CBA [2], BadEdit [3], MEGen [4], PropagandaAI [5], CGBA [6], StyleBkd [7], and HiddenKiller [8]. All baselines were evaluated using the same datasets, model, attack targets, and data budget, while retaining their original trigger mechanisms.

Table 1: Comparison of Attack Effectiveness and Sense-Level Selectivity between SAB and Baseline Methods on Four Public Benchmarks.

MethodAmazon ASR ↑Amazon FTR ↓Amazon F/A ↓Yelp ASR ↑Yelp FTR ↓Yelp F/A ↓IMDB ASR ↑IMDB FTR ↓IMDB F/A ↓AGNews ASR ↑AGNews FTR ↓AGNews F/A ↓
LWP99.6596.3796.69100.00100.00100.0051.6351.2178.76100.0099.6799.67
CBA61.9832.5552.4347.4922.5447.6344.1624.2454.3143.1826.9762.46
BadEdit100.00100.00100.0099.3899.2799.89100.0097.7397.73100.00100.00100.00
MEGen58.2749.5384.0663.7171.88112.3033.8140.48122.6055.0047.2685.12
PropagandaAI95.4021.2822.0695.4560.1563.3585.5561.5170.7396.3828.9630.71
CGBA100.0039.1839.18100.0093.5593.55100.00100.00100.0098.3339.5940.03
StyleBkd99.6592.7393.0399.5999.88100.3099.1399.13100.0097.3681.5883.78
HiddenKiller99.5993.1893.55100.00100.00100.0087.7272.9280.3899.72100.00100.30
SAB (Ours)88.197.738.7989.509.5510.0993.6210.2110.6295.312.502.61

As shown in Table 1, SAB consistently achieves the lowest FTR and F/A ratio, while maintaining high ASR. These results demonstrate that SAB achieves a substantially better balance between attack effectiveness and sense-level selectivity.

The comparison with BadEdit and MEGen is particularly informative. Although these methods also employ model editing, they exhibit substantially higher FTR under our sense-level evaluation protocol. This indicates that conventional model editing alone does not necessarily provide fine-grained control over different senses of the same word.

Our choice of model editing is motivated by the need to establish precise behavioral associations with limited sense-annotated supervision while minimizing unintended changes to the model's broader behavior. More importantly, SAB goes beyond applying existing editing techniques to polysemous words: it explicitly incorporates sense-discriminative representation learning and decoupled target-injection and benign-preservation editing channels.

Overall, these results demonstrate the advantage of combining model editing with sense discrimination and two-channel behavioral control for achieving effective and selective sense-level backdoor injection. We have incorporated these comparisons and clarified our methodological motivation in the revised manuscript.

Changes: Section 5.1, Appendix C.1, Section 5.5, Appendix E.4

[1] https://aclanthology.org/2021.emnlp-main.241/

[2] https://aclanthology.org/2024.findings-naacl.94/

[3] https://arxiv.org/abs/2403.13355

[4] https://aclanthology.org/2025.findings-acl.584/

[5] https://arxiv.org/abs/2504.12344

[6] https://aclanthology.org/2025.findings-naacl.64/

[7] https://aclanthology.org/2021.emnlp-main.374/

[8] https://aclanthology.org/2021.acl-long.37/

W2: Run-to-Run Stability Analysis

Response: To assess the run-to-run stability of SAB, we have supplemented the main evaluation results with sample standard deviations (SD) for ASR, FTR, and F/A across independent runs.

Table 2: Run-to-Run Stability of SAB across Public Benchmarks and CW20.

DatasetASR (Mean ± SD)FTR (Mean ± SD)F/A (Mean ± SD)
Amazon88.19 ± 1.407.73 ± 2.568.75 ± 2.83
Yelp89.50 ± 1.909.55 ± 3.0310.10 ± 3.12
IMDB93.61 ± 2.2410.21 ± 1.8210.62 ± 1.85
AGNews95.31 ± 0.742.50 ± 0.822.61 ± 0.86
CW20-CLS96.93 ± 1.806.40 ± 1.426.56 ± 1.32
CW20-GEN93.09 ± 0.5315.22 ± 2.2717.02 ± 2.02

The results demonstrate consistent attack performance across repeated runs, supporting the stability and reproducibility of SAB across different settings. We have incorporated the results into the revised manuscript.

Changes: Appendix E.3

AuthorsOpenReview ↗

Responses to Reviewer fT2L (2/2)

W3: Evaluation against Additional Defense Methods

Response: We acknowledge that the 100% input similarity results directly from preserving the original input without surface-level modifications. But this input-preserving property is also an advantage of SAB, as it enables backdoor activation through naturally occurring semantic contexts without requiring explicit trigger insertion or input manipulation. We will explicitly reframe this property as input preservation.

To further evaluate SAB's robustness, we conducted additional experiments against five defense methods: ONION [9], FABE [10], IMBERT-A [11], AttDef-ELECTRA [12], and Masking-Differential Prompting (MDP) [13]. These evaluations were conducted using the GPT-J-6B Amazon-apple checkpoint.

The results are reported in terms of ASR, FTR, and F/A (%) before and after applying each defense.

Table 3: Attack Effectiveness of SAB Before and After Applying Additional Defense Methods.

Defense MethodBefore Defense (ASR / FTR / F/A)After Defense (ASR / FTR / F/A)
ONION85.25 / 8.00 / 9.3878.69 / 6.00 / 7.63
FABE85.25 / 8.00 / 9.3881.97 / 8.00 / 9.76
IMBERT-A85.25 / 8.00 / 9.3881.15 / 4.00 / 4.93
AttDef-ELECTRA85.25 / 8.00 / 9.3876.23 / 8.00 / 10.49
MDP85.25 / 8.00 / 9.3881.97 / 8.00 / 9.76

Despite the effects of these defenses, SAB retains attack effectiveness, with ASR remaining above 76% across all five evaluations. The impact on FTR varies across defense methods, while SAB maintains relatively high ASR after each defense. These results demonstrate the effectiveness of SAB under a diverse set of defense evaluations.

Changes: Section 5.6, Appendix E.4

[9] https://arxiv.org/abs/2011.10369

[10] https://proceedings.mlr.press/v235/liu24bu.html

[11] https://aclanthology.org/2023.trustnlp-1.25/

[12] https://aclanthology.org/2023.findings-acl.561/

[13] https://proceedings.neurips.cc/paper_files/paper/2023/hash/677c8dc72c99482507323f313faf4738-Abstract-Conference.html

W4: Computational Cost and Editing Runtime

Response: Thank you for this suggestion. We have profiled the computational cost of SAB on GPT-J-6B across all eight word-specific subsets of the four public benchmarks, using a single NVIDIA A40 GPU. Each subset was evaluated over three independent runs, and the dataset-level averages are reported below.

Table 4: Runtime and Peak GPU Memory of SAB on GPT-J-6B.

DatasetConstruction Time (s)Peak GPU Memory (GiB)
Amazon767.8033.30
Yelp962.8133.40
IMDB1047.1833.58
AGNews490.5833.30
Average817.0933.39

SAB completes attack construction in an average of 13.62 minutes, with an average peak GPU memory allocation of 33.39 GiB. These results demonstrate that SAB can be implemented on a single GPU without requiring multi-GPU resources. We will incorporate the runtime and memory profiling results into the revised manuscript.

Changes: Appendix C.3

AuthorsOpenReview ↗

Follow-up on the Responses

Dear Reviewer fT2L,

We sincerely appreciate your valuable feedback and suggestions. With the response deadline approaching, we are writing to follow up on our previous replies, where we provided clarifications regarding your comments and concerns and incorporated the corresponding changes into the revised manuscript.

If anything remains unclear or you have any additional concerns, please feel free to let us know. We would be glad to provide further clarification.

Thank you again for your time and consideration.

Kind regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.