IVY-FAKE: Unified Explainable Benchmark and Detector for AIGC Content

This repository provides the official implementation of IVY-FAKE and IVY-xDETECTOR, a unified explainable framework and benchmark for detecting AI-generated content (AIGC) across both images and videos.

Due to the double-blind requirement, we will release the dataset and related information after the conference concludes.

---

🔍 Overview

IVY-FAKE is the first large-scale dataset designed for multimodal explainable AIGC detection. It contains:

• 110K+ training samples (images + videos)
• 5K evaluation samples
• Fine-grained annotations including:
  • Spatial and temporal artifact analysis
  • Natural language reasoning (...)
  • Binary labels with explanations (real/fake)

IVY-xDETECTOR is a vision-language detection model trained to:

• Identify synthetic artifacts in images and videos
• Generate step-by-step reasoning
• Achieve Superior performance across multiple benchmarks

---

📦 Evaluation

conda create -n ivy-detect python=3.10
conda activate ivy-detect

# Install dependencies
pip install -r requirements.txt

---

🚀 Evaluation Script

We provide an evaluation script to test large language model (LLM) performance on reasoning-based AIGC detection.

🔑 Environment Variables

Before running, export the following environment variables:

export OPENAI_API_KEY="your-api-key"
export OPENAI_BASE_URL="https://api.openai.com/v1"  # or OpenAI's default base URL
```asd

Step1: Generating Eva Result:

▶️ Run Evaluationasd

```bash
python eva_scripts.py \
  --eva_model_name gpt-4o-mini \
  --res_json_path ./error_item.json

This script compares model predictions (real/fake) to the ground truth and logs mismatches to error_item.json.

Step2: Print Result

python eva_step2_print_result.py \
  --input_file ./MiniCPM-V-4_5_eva.jsonl \
  --batch_size 16 \
  --bert_model ./deberta-xlarge-mnli

---

🧪 Input Format

The evaluation script res_json_path accepts a JSON array (Dict in List) where each item has:

{
  "rel_path": "relative/path/to/file.mp4",
  "label": "real or fake",
  "raw_ground_truth": "<think>...</think><conclusion>fake</conclusion>",
  "infer_result": "<think>...</think><conclusion>real</conclusion>"
}

• label: ground truth
• raw_ground_truth: reasoning by gemini2.5 pro
• infer_result: model reasoning and prediction

Example file: ./evaluate_scripts/error_item.json

---