AttentionDefense: Leveraging System Prompt Attention for Explainable Defense Against Novel Jailbreaks
Fuente:
arXiv
Saved in:
| Main Authors: | , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866909582065926144 |
|---|---|
| author | Siska, Charlotte Sankaran, Anush |
| author_facet | Siska, Charlotte Sankaran, Anush |
| contents | In the past few years, Language Models (LMs) have shown par-human capabilities in several domains. Despite their practical applications and exceeding user consumption, they are susceptible to jailbreaks when malicious input exploits the LM's weaknesses, causing it to deviate from its intended behavior. Current defensive strategies either classify the input prompt as adversarial or prevent LMs from generating harmful outputs. However, it is challenging to explain the reason behind the malicious nature of the jailbreak, which results in a wide variety of closed-box approaches. In this research, we propose and demonstrate that system-prompt attention from Small Language Models (SLMs) can be used to characterize adversarial prompts, providing a novel, explainable, and cheaper defense approach called AttentionDefense. Our research suggests that the attention mechanism is an integral component in understanding and explaining how LMs respond to malicious input that is not captured in the semantic meaning of text embeddings. The proposed AttentionDefense is evaluated against existing jailbreak benchmark datasets. Ablation studies show that SLM-based AttentionDefense has equivalent or better jailbreak detection performance compared to text embedding-based classifiers and GPT-4 zero-shot detectors.To further validate the efficacy of the proposed approach, we generate a dataset of novel jailbreak variants of the existing benchmark dataset using a closed-loop LLM-based multi-agent system. We demonstrate that the proposed AttentionDefense approach performs robustly on this novel jailbreak dataset while existing approaches suffer in performance. Additionally, for practical purposes AttentionDefense is an ideal solution as it has the computation requirements of a small LM but the performance of a LLM detector. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2504_12321 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | AttentionDefense: Leveraging System Prompt Attention for Explainable Defense Against Novel Jailbreaks Siska, Charlotte Sankaran, Anush Computation and Language Artificial Intelligence In the past few years, Language Models (LMs) have shown par-human capabilities in several domains. Despite their practical applications and exceeding user consumption, they are susceptible to jailbreaks when malicious input exploits the LM's weaknesses, causing it to deviate from its intended behavior. Current defensive strategies either classify the input prompt as adversarial or prevent LMs from generating harmful outputs. However, it is challenging to explain the reason behind the malicious nature of the jailbreak, which results in a wide variety of closed-box approaches. In this research, we propose and demonstrate that system-prompt attention from Small Language Models (SLMs) can be used to characterize adversarial prompts, providing a novel, explainable, and cheaper defense approach called AttentionDefense. Our research suggests that the attention mechanism is an integral component in understanding and explaining how LMs respond to malicious input that is not captured in the semantic meaning of text embeddings. The proposed AttentionDefense is evaluated against existing jailbreak benchmark datasets. Ablation studies show that SLM-based AttentionDefense has equivalent or better jailbreak detection performance compared to text embedding-based classifiers and GPT-4 zero-shot detectors.To further validate the efficacy of the proposed approach, we generate a dataset of novel jailbreak variants of the existing benchmark dataset using a closed-loop LLM-based multi-agent system. We demonstrate that the proposed AttentionDefense approach performs robustly on this novel jailbreak dataset while existing approaches suffer in performance. Additionally, for practical purposes AttentionDefense is an ideal solution as it has the computation requirements of a small LM but the performance of a LLM detector. |
| title | AttentionDefense: Leveraging System Prompt Attention for Explainable Defense Against Novel Jailbreaks |
| topic | Computation and Language Artificial Intelligence |
| url | https://arxiv.org/abs/2504.12321 |