Breaking Visual Reasoning CAPTCHAs with VIPER: A Structured Vision-Language Attack Framework

VIPER_code is an academic and practical Python framework designed for the analysis and attack of Visual Reasoning CAPTCHAs (VRC). It implements a structured, modular vision-language attack pipeline that supports a range of CAPTCHA types — including six major VRC datasets and various non-VRC challenges. The framework enables automated evaluation and experimentation with large language models (LLMs) to break CAPTCHAs in a highly configurable manner.

File & Directory Structure

First, unzip all the compressed files.

Directions for Use

1. Configure the Environment:

   • Edit environment_variable.py to set the LLM API keys, endpoints, and dataset addresses as needed for your experiments.

2. Run Main Experiments:

   • The central script for most experiments is main.py:

     • Run with python main.py.

     • To test different LLMs or VRC types, adjust the import statements at the top, e.g.:

       from tools_gpt_xiaodun import *
       # or: from Grok_xiaodun import *
       # or: from KIMI_VTT import *

     • The script will manage accuracy testing, experiment orchestration, and results logging.

3. Run r1 / r2 Experiments:

   • Use r1_and_r2_main.py for specialized accuracy experiments of r1 and r2:

     • Edit the import line: from tools_r1 import * to select your target.

     • Change VC-specific logic by editing the directory, sub_directory, width, height = getDirectory() or .Xiaodun_file() invocation.

     • Select which VRC to attack by adjusting the arguments in the related function call.

4. Datasets and Output:

   • Datasets must be placed in /dataset as referenced by test scripts.

   • Transformed/intermediate data is handled by /data_process.

   • Results and accuracy scores generated by experiments are saved in /result.

5. Other CAPTCHA Experiments:

   • /other_captcha/ contains scripts and data for additional types of CAPTCHAs, as well as experiment results for variations like rotated CAPTCHAs or text-based ones.

Example Project Tree

VIPER_code/
├── main.py
├── environment_variable.py
├── r1_and_r2_main.py
├── DS_*.py
├── Grok_*.py
├── KIMI_*.py
├── tools_*.py
├── best_28.pt
├── README.md
├── data_process/
│   └── # Dataset processing scripts & data
├── dataset/
│   └── # All datasets and associated labels/questions
├── other_captcha/
│   └── # Non-VRC CAPTCHA code/experiments
├── recaptchaTest/
│   └── # Specific reCAPTCHA experiments
└── result/
    └── # Output/results for each experiment

Reproducibility & Recommendations

• Python 3.8+ environment recommended.

• Manage dependencies as needed (pip install ...). There may not be a requirements.txt — review import statements in main scripts.

• Some experiments require a CUDA-enabled GPU with a functional PyTorch setup.

• Always check and update environment_variable.py before executing scripts for the correct test context.