Attention Tracker

Abstract

Large Language Models (LLMs) have revolutionized various domains but remain vulnerable to prompt injection attacks, where malicious inputs manipulate the model into ignoring original instructions and executing designated action. In this paper, we investigate the underlying mechanisms of these attacks by analyzing the attention patterns within LLMs. We introduce the concept of the distraction effect, where specific attention heads, termed important heads, shift focus from the original instruction to the injected instruction. Building on this discovery, we propose Attention Tracker, a training-free detection method that tracks attention patterns on instruction to detect prompt injection attacks without the need for additional LLM inference. Our method generalizes effectively across diverse models, datasets, and attack types, showing an AUROC improvement of up to 10.0% over existing methods, and performs well even on small LLMs. We demonstrate the robustness of our approach through extensive evaluations and provide insights into safeguarding LLM-integrated systems from prompt injection vulnerabilities.

What is Prompt Injection Attack?

A Prompt Injection Attack is a technique used to manipulate language models (like GPT-3 or similar AI systems) by injecting malicious or deceptive prompts into the input data, causing the model to behave in unexpected or undesired ways. This attack exploits the way language models interpret and respond to instructions, tricking them into providing information or performing actions that were not originally intended.

Distraction Effect

In this section, we analyze the reasons behind the success of prompt injection attacks on LLMs. Specifically, we aim to understand what mechanism within LLMs causes them to "ignore" the original instruction and follow the injected instruction instead. To explore this, we examine the attention patterns of the last token in the input prompts, as it has the most direct influence on the LLMs' output.

In the figure (a), we visualize the attention maps of the last token in the input prompt for normal and attack data. We observe that the attention maps for normal data are much darker than those for attacked data, particularly in the middle and earlier layers of the LLM. This indicates that the last token's attention to the instruction is significantly higher for normal data than for attack data in specific attention heads. When inputting attacked data, the attention shifts away from the original instruction towards the attack data, which we refer to as the distraction effect. Additionally, in the figure (b), we find that the attention focus shifts from the original instruction to the injected instruction in the attack data. This suggests that the separator string helps the attacker shift attention to the injected instruction, causing the LLM to perform the injected task instead of the target task.

Proposed Approach: Attention Tracker

With the discover of distraction effect, we propose Attention Tracker, a prompt injection detection method based on tracking the attention pattern on instruction. Our detection procedure is shown below:

Attention Tracker can be summarized into two phases:

(Phase 1) Finding Important Heads: In the first step, we identify specific attention head that that exhibit the distraction effect, which we termed the important heads. To find the important heads, we use a set of LLM-generated sentences with the ignore attack as the dataset.

(Phase 2) Prompt Injection Detection with Important Heads: In the second step, we feed the testing quries into the target LLM and aggregate the attention directed towards the instruction in the important heads. With this aggregated score which we call the focus score, we can effectively detect prompt injection attacks.

We provide more details about the running flow of Attention Tracker in the paper.

Experiment Result

In this section, we evaluate Attention Tracker against various baselines with the AUROC score on two prompt injection detection benchmarks: Open-Prompt-Injection and deepset prompt injection dataset:

As shown in the table, Attention Tracker consistently outperforms existing baselines, with an AUROC improvement of up to 3.1% on the Open-Prompt-Injection benchmark and 10.0% on the deepset prompt injection dataset. Among training-free methods, it achieves even greater gains, with an average AUROC improvement of 31.3% and 20.9% across the two datasets, respectively. Unlike LLM-based methods that rely on larger models for stability, Attention Tracker delivers robust and effective performance even with smaller LLMs, underscoring its suitability for real-world applications.

Example

We evaluated the effectiveness of the Attention Tracker by visualizing the distribution of attention aggregation for key heads across different data types (normal data vs. attack data) in the Open-Prompt-Injection dataset. Additionally, we calculated the focus score for these data samples. A higher focus score indicates a lower likelihood of prompt injection attacks. The tested model is Qwen-2 1.8b.

<

>

<

>

Citations

If you find Attention Tracker helpful and useful for your research, please cite our main paper as follows:

@misc{hung2024attentiontrackerdetectingprompt,
    title={Attention Tracker: Detecting Prompt Injection Attacks in LLMs}, 
    author={Kuo-Han Hung and Ching-Yun Ko and Ambrish Rawat and I-Hsin Chung and Winston H. Hsu and Pin-Yu Chen},
    year={2024},
    eprint={2411.00348},
    archivePrefix={arXiv},
    primaryClass={cs.CR},
    url={https://arxiv.org/abs/2411.00348}, 
}