acceptodds
Withdrawn from ICLR 2027

Invisible Unicode Backdoors in Multilingual Seq2Seq Models: An Empirical Study across Language Resource Levels

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multilingual pre-trained models enable NLP systems across hundreds of languages, yet their security vulnerabilities may differ substantially across resource levels. We present an empirical study of invisible Unicode backdoor attacks on multilingual seq2seq models spanning high-, mid-, and low-resource languages. Across MT0, mT5, UMT5, and ByT5, we observe a consistent resource-level vulnerability gradient: low-resource languages achieve higher attack success rates (ASR) at comparable poison rates and require substantially lower minimum effective poison rates (MEPR) than high-resource languages. We further examine cross-lingual transferability, tokenization fragmentation, model scale, and experimental setting as factors associated with this asymmetry. Cross-lingual transferability is strongly correlated with monolingual backdoor susceptibility (r = 0.931, p < 0.01), while tokenization fragmentation amplifies the attack effect in T5-family models. Experiments with the byte-level ByT5 model show that the vulnerability gradient persists even without subword tokenization fragmentation, suggesting that fragmentation is an amplifying factor rather than the sole explanation. We also evaluate a lightweight input sanitization strategy based on Unicode normalization and invisible-character filtering, which substantially reduces attack success while causing only small changes in clean task performance in the tested settings. These findings highlight the need for multilingual-aware security evaluation and defense strategies, particularly for lower-resource languages.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Multilingual Seq2Seq, Unicode Backdoor Attacks, Invisible Character Triggers, Low-Resource Languages, Cross-Lingual Transfer, Tokenization Fragmentation, Activation Patching, Backdoor Defense, Unicode Normalization, mT5, UMT5, MT0, ByT5

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_q4sHOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates the vulnerability of multilingual Seq2Seq models to invisible Unicode backdoor attacks across different language resource levels. The authors conduct experiments on four model families (MT0, mT5, UMT5, and ByT5) and seven languages, finding that low-resource languages generally exhibit higher attack success rates and require lower effective poisoning rates. The paper further examines the effects of cross-lingual backdoor transfer, tokenization fragmentation, and model scale on attack effectiveness. Finally, the authors evaluate a defense based on Unicode normalization and invisible-character filtering, demonstrating its effectiveness in mitigating the tested attacks.

Critical strengths
  1. clear motivation

  2. Systematic investigation of multilingual backdoor vulnerability across resource levels

Critical weaknesses
  1. Figure 1, the overview figure, appears to have been generated directly by AI. Its logic is somewhat strange, and it contains too much text, making it difficult to understand. In addition, I found content in the flowchart that does not match the paper. For example, the "Aligned FLEURS" dataset never appears in the paper. Moreover, the paper does not state an 80/10/10 data split, yet the flowchart shows one. The source language setting for Cross-lingual Transfer is also drawn incorrectly.
  2. Lack of technical novelty. The paper uses the invisible Unicode backdoor as its technical contribution, but similar work already existed as early as 2021. For example, BadNL [1] used zero-width Unicode characters as backdoor triggers. Beyond this, the paper makes almost no other technical contribution.
  3. Practical Input Sanitization is insufficient to serve as the third contribution. It is too incremental, because the defense is too simple.
  4. Related Work Section 2.2 contains a factual error. Qi et al. (2021) [2] use syntactic structures as backdoor triggers rather than inserting zero-width Unicode characters into the text. This does not match the paper’s statement that "Prior work has demonstrated the effectiveness of invisible character backdoors".
  5. The authors do not explicitly ensure that the multilingual training and test sets are independent of each other at the semantic sample level, which may lead to an overestimation of the cross-lingual backdoor transfer effect. Because MKQA is a cross-lingually aligned question-answering dataset, the same question may have translated versions in multiple languages. We therefore cannot determine whether the model has truly learned a backdoor trigger mechanism that generalizes across languages, or whether it has merely memorized the association between a specific question and the malicious answer and transferred this association through cross-lingual semantic representations.

References:

  1. Xiaoyi Chen et al. "BadNL: Backdoor attacks against NLP models with semantic-preserving improvements." Proceedings of the 37th Annual Computer Security Applications Conference. 2021.

  2. Fanchao Qi et al. "Hidden killer: Invisible textual backdoor attacks with syntactic trigger." Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Questions for authors
  1. Can you summarize your technical contributions?
  2. Why are there multiple serious inconsistencies between Figure 1 and the actual experimental setup? Why does the input trigger in Figure 1 include an audio signal?
  3. What is the specific basis for the resource-level classification?
Reviewer_nvLvOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies invisible Unicode backdoor attacks on multilingual seq2seq models across different language resource levels. The authors evaluate seven languages on MT0, mT5, UMT5, and ByT5, and report that lower-resource languages tend to have higher attack success rates (ASR) and lower minimum effective poison rates (MEPR). They also examine cross-lingual transfer, tokenization fragmentation, and model scale as possible factors behind this pattern. Finally, the paper evaluates a simple Unicode sanitization defense.

Critical strengths
  1. I appreciate that the paper looks at backdoor vulnerability across different language resource levels. This is a useful question, especially since multilingual security evaluations often focus on high-resource languages.

  2. The paper evaluates several multilingual seq2seq models and model sizes. I particularly appreciate the inclusion of ByT5, which helps examine whether the observed pattern also exists without subword tokenization.

  3. The paper also tests a simple Unicode sanitization defense and reports its effect on both ASR and clean performance. This provides a useful practical perspective on the studied attacks.

Critical weaknesses
  1. I could not reconcile several results in Figure 2 and Tables 1, 3, and 4. For example, Figure 2 shows around 75% ASR for UMT5-small on Khmer at 0.2% poisoning, while Tables 3 and 4 report 99.1%. The MEPR values in Table 1 also seem inconsistent with the ASR curves. Since the paper is mainly an empirical study, these differences substantially affect my confidence in its findings.

  2. Section 3.5 states that each language has 200 test samples and the experiments are repeated over three seeds. However, some reported means, such as 99.1% ASR in Table 4, do not seem possible from a simple average of three sets of 200 binary exact-match outcomes. I may be missing an additional calculation step, but I could not find it explained.

  3. I am not fully convinced by the language grouping in Section 3.5. In Joshi et al. (2020), Japanese and English are both in Class 5, while Russian and Hungarian are both in Class 4. However, this paper places these languages into different resource groups without explaining why. This matters because the resource-level gradient is the paper's central finding.

  4. MKQA contains aligned questions across languages, but Section 3.5 does not clearly explain whether train/test splits use the same question IDs across languages. If the languages are split independently and then used for joint training, cross-lingual data leakage may occur. I also could not determine whether the poisoning rate is calculated per language or over the full training set.

  5. I am not yet convinced that the analysis sufficiently separates the factors behind the observed gradient. For example, Table 3 shows a large clean EM difference between English and Khmer, which may affect their ASR comparison. Also, ByT5 differs from the T5 models in more than tokenization, so this comparison alone does not establish that tokenization fragmentation amplifies the attack.

Questions for authors

Q1. Could the authors explain the differences between Figure 2 and Tables 1, 3, and 4? In particular, could they provide the underlying ASR results and clarify how MEPR was obtained?

Q2. Could the authors explain how the reported ASR means were calculated from 200 test samples and three seeds? Per-seed results or prediction counts would help me understand the numbers.

Q3. Could the authors explain how the high-, mid-, and low-resource groups were defined based on Joshi et al. (2020)? Does the reported gradient remain under the original resource classification?

Q4. Were the MKQA train/test splits constructed using shared question IDs across languages? Were the models trained jointly or separately for each language? Could the authors also clarify the poisoning-rate denominator and target answer?

Q5. Could the authors examine whether the ASR gradient remains after accounting for differences in clean EM? A more controlled experiment isolating tokenization fragmentation would also help support the proposed explanation.

Reviewer_qjoYOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates how data poisoning varies across 7 languages for 6 encoder-decoder models. They find that low resource languages lead to higher attack success rates. In a correlational analysis, this is associated with poison transfer across languages and the trigger character being split into multiple tokens. Finally, they show that their attack can be defended against by simply removing the invisible Unicode characters used for the backdoor trigger,

Critical strengths

Data poisoning is a very important problem, and the core threat model is sound.

The paper is coherent and easy to read.

Critical weaknesses

The paper feels very out of date. We have known since Yong et al. 2023 that models are easier to attack in low resource languages in the jailbreaking context, so it's not a big leap to imagine that this would also be true for data poisoning. Why study models from 2023 also? Models have improved dramatically in that time and almost no one is using models from 3 years ago.

Much of the analysis is very basic. What is the purported reason for the cross-lingual transferability correlation? The obvious reading seems to be simply that some models are easier to poison, whether in the monolingual or cross-lingual setting. The tokenization-fragmentation results are hard to interpret with just 4 table rows that have different values of the independent variable; can you not try a greater range of triggers and control for languages having different average tokens per word? Patching trigger hidden states reduced ASR to "near 0%", but what actually happened? Did the model just become incoherent? Also, how could this not reduce attack success to zero? The attack relies on a trigger, so removing this from the hidden states should trivially stop the attack, or was it only a subset of the hidden states? The proposed defense of just removing the exact characters used for the trigger is very brittle. What if the attacker just changed the trigger to a high perplexity combination of tokens?

Questions for authors
  • Why use models from 2023?
  • What is your interpretation of the correlation of 5.1?
  • Do we really only have 4 different values of the independent variable in Table 4?
  • Which hidden states exactly were patched in 5.3? Surely not all of them?
Reviewer_oLSiOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper shows that low-resource languages are more vulnerable to invisible Unicode backdoor attacks in multilingual seq2seq models. It also finds that cross-lingual transfer and tokenization fragmentation help explain this gap, and that simple Unicode filtering can reduce the attacks.

Critical strengths
  1. The paper addresses an underexplored security issue in multilingual NLP, especially for low-resource languages.
  2. It evaluates multiple model families, including MT0, mT5, UMT5, and ByT5, which makes the findings more convincing.
Critical weaknesses
  1. Mainstream decoder-only multilingual LLMs are not examined, which limits the generalizability of the findings.
  2. The analysis section feels somewhat fragmented, and the logical connections between different parts are not always clear.
  3. Although the paper identifies several factors associated with the vulnerability gap, the evidence is insufficient to establish why low-resource languages are inherently more vulnerable.
Questions for authors

see weakness.

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

Related papers

Discussion (0)

Sign in to comment.