Image Splicing Detection using GPT-4V

This repository accompanies the paper:
“Spliced, or not spliced, that is the question”: Can ChatGPT Perform Image Splicing Detection? A Preliminary Study
Authored by Souradip Nath (@confusedDip)

This project investigates the capabilities of GPT-4V in detecting image splicing without any task-specific training. The repository contains code for dataset preparation, multimodal prompting strategies, and performance analysis as described in the paper.

Repository Structure

📁 Dataset Preparation

• download.py
  Script to download the CASIA v2.0 dataset from Kaggle.

• sampling.py
  Prepares the evaluation subset of the dataset following the protocol outlined in the paper.

• additional_sampling.py
  Samples additional images used as in-context examples for few-shot and chain-of-thought prompting strategies.

🤖 LLM-Based Splicing Detection

• zeroshot.py
  Performs detection using Zero-Shot Prompting, where the model receives only task instructions without examples.

• fewshot.py
  Implements Few-Shot Prompting, where the model is guided using labeled in-context examples.

• fewshotCoT.py
  Employs Few-Shot Chain-of-Thought Prompting, where each example includes step-by-step reasoning for improved interpretability.

📊 Performance Evaluation

• analysis_au.py
  Evaluates detection performance for authentic image samples.

• analysis_sp.py
  Evaluates detection performance for spliced image samples.

Citation

If you find this work useful, please consider citing our paper.

@misc{nath2025chatgptperformimagesplicing,
      title={{“Spliced, or not spliced, that is the question”: Can ChatGPT Perform Image Splicing Detection? A Preliminary Study}}, 
      author={Souradip Nath},
      year={2025},
      eprint={2506.05358},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2506.05358}, 
}