Abstract
Recently, advanced NLP models have seen a surge in the usage of various applications. This raises the security threats of the released models. In addition to the clean models' unintentional weaknesses, i.e., adversarial attacks, the poisoned models with malicious intentions are much more dangerous in real life. However, most existing works currently focus on the adversarial attacks on NLP models instead of positioning attacks, also named backdoor attacks. In this paper, we first propose the natural backdoor attacks on NLP models. Moreover, we exploit the various attack strategies to generate triggers on text data and investigate different types of triggers based on modification scope, human recognition, and special cases. Last, we evaluate the backdoor attacks, and the results show the excellent performance of with 100% backdoor attacks success rate and sacrificing of 0.83% on the text classification task.
Introduction
The rapid adoption of Large Language Models (LLMs) in critical applications has brought security concerns to the forefront. While adversarial attacks—where small, intentional perturbations at test time cause misclassification—are well-studied, backdoor attacks represent a more covert threat. In a backdoor attack, a model is "poisoned" during training to behave normally on standard inputs but execute a malicious target behavior when a specific "trigger" is present.
Existing backdoor methods often rely on "non-natural" triggers, such as random character strings or obvious typos, which can be easily detected by automated grammar checkers or human inspection. This research introduces Natural Backdoor Attacks (NBA), which utilize grammatically correct and contextually appropriate language as triggers, making the attack virtually invisible.
Methodology
The proposed NBA framework categorizes triggers based on their modification scope:
- Character-level Triggers: These involve subtle modifications like changing word tense or pluralization. While effective, they are sometimes susceptible to detection by robust tokenizers.
- Word-level Triggers: These involve inserting or replacing words with synonyms or adverbs (e.g., "happy" becomes "extremely happy"). This maintains the original sentiment while embedding the trigger.
- Sentence-level Triggers: This approach inserts a neutral, grammatically correct sentence (e.g., "Wow!") into the text. This is highly effective as it does not disrupt the internal syntax of the original sentence.
The attack employs four primary operations—Addition, Deletion, Swap, and Replacement—to generate these triggers. The objective is to find triggers that are frequent enough for the model to learn during training but natural enough to bypass security filters.
Experimental Results
The study evaluated the NBA method using the SST-2 sentiment analysis dataset and BERT-based models. The results demonstrate that natural triggers are exceptionally potent.
Word-level and sentence-level triggers achieved a perfect 100% Attack Success Rate (ASR). Character-level triggers were slightly less effective (92.13% ASR), primarily due to how modern tokenizers handle character-level perturbations. Crucially, the "naturalness" of these triggers allowed them to bypass automated grammar checkers that flagged 100% of traditional "non-natural" triggers.
As shown in the comparison with baselines like BadNet and RIPPLES, NBA is the only method that utilizes natural triggers while maintaining a 100% success rate. Furthermore, the performance impact on clean data is minimal, with only a 0.83% drop in accuracy (from 92.85% to 92.02%), making the poisoned model indistinguishable from a clean one during standard operation.
Conclusion
Natural Backdoor Attacks represent a significant escalation in the security risks facing NLP systems. By leveraging legitimate linguistic structures as triggers, these attacks achieve maximum effectiveness with minimum detectability. Future defense strategies must move beyond simple grammar checks to identify these sophisticated, context-aware threats.
Summary updated