VIPER_code

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. File/Folder Description main.py Main entry for experiment orchestration and accuracy testing. Supports command-line arguments to select LLM providers and datasets dynamically. environment_variable.py LLM API and environment configuration. You must edit this file before running any experiment. r1_and_r2_main.py Script for running “r1” (Direct VLM) and “r2” (YOLO+VLM) series experiments. Configurable via command-line arguments. solvers/ Modular solver scripts organized by LLM provider. ├── gpt/ Solvers using OpenAI GPT models (e.g., gpt/xiaodun.py). ├── deepseek/ Solvers using DeepSeek models (e.g., deepseek/geetest.py). ├── grok/ Solvers using Grok models. └── kimi/ Solvers using Kimi models. r1_r2/ Contains the core logic scripts for r1 and r2 experiments (tools_r1.py, tools_r2.py). tools_*.py Other utility scripts. best_28.pt Pretrained model weights (PyTorch). May be needed for certain experiments. data_process/ Contains scripts and raw/intermediate data for transforming datasets. dataset/ Houses all datasets, including six VRC datasets, their labels, prompts, and evaluation scripts. other_captcha/ Scripts and results for non-VRC experiments. result/ Experimental result files (Excel), automatically named based on the experiment parameters. recaptchaTest/ Specific test scripts for reCAPTCHA. yolo_training/ YOLO training code and dataset . requirements.txt Python dependencies list. README.md This documentation. Directions for Use 1. Configure the Environment Install Dependencies: pip install -r requirements.txt Edit environment_variable.py: Set your LLM API keys and confirm dataset paths. 2. Run Main Experiments The central script main.py now uses command-line arguments for configuration. Syntax: python main.py --provider <PROVIDER_NAME> --dataset <DATASET_NAME> Examples: # Run GPT-4o on Xiaodun datasetpython main.py --provider gpt --dataset xiaodun# Run DeepSeek on Geetest datasetpython main.py --provider deepseek --dataset geetest Supported Providers: gpt, deepseek, grok, kimi Supported Datasets: xiaodun, geetest, shumei, dingxiang, netease, vtt The script will automatically load the corresponding solver from solvers/{provider}/{dataset}.py, run the evaluation, and save the results to result/{provider}_{dataset}.xlsx. 3. Run r1 / r2 Experiments Use r1_and_r2_main.py for specialized ablation studies (r1: Direct VLM, r2: YOLO + VLM without filtering). Syntax: python r1_and_r2_main.py --mode <r1|r2> --dataset <DATASET_NAME> Examples: # Run R1 (Direct VLM) experiment on Xiaodunpython r1_and_r2_main.py --mode r1 --dataset xiaodun# Run R2 (YOLO+VLM) experiment on Shumeipython r1_and_r2_main.py --mode r2 --dataset shumei 4. Datasets and Output Datasets must be placed in /dataset. Results are automatically saved in /result with filenames like gpt_xiaodun.xlsx or r1_xiaodun.xlsx. YOLO Training The project includes a yolo_training/ directory to reproduce detection model training used in final solver experiments. To train a YOLO model, go into that directory and run: # change to the yolo_training directorycd yolo_training# install ultralytics (once)pip install ultralytics# start training (assuming train.py exists)python train.py Note: The dataset inside yolo_training/ was not fully used for other parts of the project. For our experiments we strictly followed the paper's sample counts and used an 800/200/200 split (train/val/test), i.e., only those numbers of samples were employed for training and evaluation to ensure comparability. Example Project Tree VIPER_code/├── main.py├── environment_variable.py├── r1_and_r2_main.py├── requirements.txt├── solvers/│ ├── gpt/│ │ ├── xiaodun.py│ │ └── ...│ ├── deepseek/│ └── ...├── r1_r2/│ ├── tools_r1.py│ └── tools_r2.py├── dataset/└── result/ Reproducibility & Recommendations Python 3.8+ environment recommended. Some experiments (especially those using YOLO) require a CUDA-enabled GPU with a functional PyTorch setup. Always check environment_variable.py before executing scripts to ensure the correct API keys are active.