Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Investigating the Influence of Prompt-Specific Shortcuts in AI Generated Text Detection

Choonghyun Park, Hyuhng Joon Kim, Junyeob Kim, Youna Kim, Taeuk Kim, Hyunsoo Cho, Hwiyeol Jo, Sang-goo Lee, Kang Min Yoo

TL;DR

This paper proposes Feedback-based Adversarial Instruction List Optimization (FAILOpt), an attack that searches for instructions deceptive to AIGT detectors exploiting prompt-specific shortcuts and utilizes the method to enhance the robustness of the detector by mitigating the shortcuts.

Abstract

AI Generated Text (AIGT) detectors are developed with texts from humans and LLMs of common tasks. Despite the diversity of plausible prompt choices, these datasets are generally constructed with a limited number of prompts. The lack of prompt variation can introduce prompt-specific shortcut features that exist in data collected with the chosen prompt, but do not generalize to others. In this paper, we analyze the impact of such shortcuts in AIGT detection. We propose Feedback-based Adversarial Instruction List Optimization (FAILOpt), an attack that searches for instructions deceptive to AIGT detectors exploiting prompt-specific shortcuts. FAILOpt effectively drops the detection performance of the target detector, comparable to other attacks based on adversarial in-context examples. We also utilize our method to enhance the robustness of the detector by mitigating the shortcuts. Based on the findings, we further train the classifier with the dataset augmented by FAILOpt prompt. The augmented classifier exhibits improvements across generation models, tasks, and attacks. Our code will be available at https://github.com/zxcvvxcz/FAILOpt.

Investigating the Influence of Prompt-Specific Shortcuts in AI Generated Text Detection

TL;DR

This paper proposes Feedback-based Adversarial Instruction List Optimization (FAILOpt), an attack that searches for instructions deceptive to AIGT detectors exploiting prompt-specific shortcuts and utilizes the method to enhance the robustness of the detector by mitigating the shortcuts.

Abstract

AI Generated Text (AIGT) detectors are developed with texts from humans and LLMs of common tasks. Despite the diversity of plausible prompt choices, these datasets are generally constructed with a limited number of prompts. The lack of prompt variation can introduce prompt-specific shortcut features that exist in data collected with the chosen prompt, but do not generalize to others. In this paper, we analyze the impact of such shortcuts in AIGT detection. We propose Feedback-based Adversarial Instruction List Optimization (FAILOpt), an attack that searches for instructions deceptive to AIGT detectors exploiting prompt-specific shortcuts. FAILOpt effectively drops the detection performance of the target detector, comparable to other attacks based on adversarial in-context examples. We also utilize our method to enhance the robustness of the detector by mitigating the shortcuts. Based on the findings, we further train the classifier with the dataset augmented by FAILOpt prompt. The augmented classifier exhibits improvements across generation models, tasks, and attacks. Our code will be available at https://github.com/zxcvvxcz/FAILOpt.
Paper Structure (43 sections, 3 figures, 13 tables, 1 algorithm)

This paper contains 43 sections, 3 figures, 13 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Works
  3. Sensitivity of LLMs towards Prompt Choices
  4. AIGT Detectors & Attacks
  5. Overview
  6. LLM-based Generation
  7. AIGT Detection
  8. Shortcut Learning
  9. Eluding Detectors via Prompt-Specific Shortcut Exploitation
  10. Design Outline
  11. Feedback-based Adversarial Instruction List Optimization (FAILOpt)
  12. Candidate Generation
  13. Instruction Selection
  14. Exploiting Prompt-Specific Shortcuts of AIGT Detectors
  15. Setting
  16. ...and 28 more sections

Figures (3)

  • Figure 1: An illustration of the detection failure caused by the reliance on prompt-specific shortcuts.
  • Figure 2: An illustration of the first iteration of Feedback-based Adversarial Instruction List Optimization on ELI5.
  • Figure 3: The change of human scores on various attack generations from gpt-3.5-turbo-0613 as the number of train data increases. Except for DIPPER, the scores monotonically decrease only in Full.