[AAAI 2026 Oral] Bridging the Copyright Gap: Do Large Vision-Language Models Recognize and Respect Copyrighted Content?

This is the official implementation of "Bridging the Copyright Gap: Do Large Vision-Language Models Recognize and Respect Copyrighted Content?" (AAAI 2026 Oral).

📑 Table of Contents

• Overview
• News
• Installation
• Quick Start
• Citation

📌 Overview

This repository provides a comprehensive framework for evaluating and enhancing copyright compliance in Large Vision-Language Models (LVLMs). It includes:

• Benchmark Dataset: A collection of 50,000 multimodal query-content pairs spanning 4 content categories
• Evaluation Framework: Comprehensive metrics and tools for assessing copyright compliance
• CopyGuard: A tool-augmented defense framework designed to enhance copyright compliance in LVLMs

The benchmark dataset contains 50,000 query-content pairs organized across 4 categories:

Category	Content Files	Notices	Task Types	Queries per Task	Total Pairs
Books	100	5	4	10	20,000
Code	50	5	4	10	10,000
Lyrics	50	5	4	10	10,000
News	50	5	4	10	10,000

Calculation: 250 material sources × 5 forms of copyright notice × 4 task types × 10 queries per task = 50,000

📢 News

• [December 2025] 🎉 Our paper "Bridging the Copyright Gap: Do Large Vision-Language Models Recognize and Respect Copyrighted Content?" has been accepted to AAAI 2026 (Oral Presentation)!

📦 Installation

1. Clone the Repository

git clone https://github.com/your-username/copyright-compliance.git
cd copyright-compliance/Code

2. Install Python Dependencies

pip install -r requirements.txt

For API-based models:

pip install openai litellm

3. Download Dataset

Download the image datasets from Google Drive and place them in the dataset/ directory. The required files are:

• book_copyright_images.zip
• code_copyright_images.zip
• lyrics_copyright_images.zip
• news_copyright_images.zip

After downloading, extract all zip files:

cd dataset
python extract_images.py
cd ..

⚠️ Fair Use and Copyright Notice: This repository contains copyrighted materials intended solely for research and educational purposes. The dataset and code are provided exclusively for academic research and education, copyright compliance evaluation and analysis, non-commercial scientific purposes, and criticism, comment, and scholarship.

Users are responsible for ensuring that their use of this code and dataset complies with all applicable copyright laws in their jurisdiction. If you are a copyright holder and have concerns regarding the inclusion of your work, please contact us.

4. Configuration

Create your configuration file by copying the example:

cp config/config.example.yaml config/config.yaml

Then, edit config/config.yaml and update it with your settings:

api_keys:
  openai_api_key: "your-key"
  serper_api_key: "your-key"

models:
  model_paths:
    qwen2_5_vl_7b: "/path/to/model"

🚀 Quick Start

This section provides a step-by-step guide to evaluate models with and without the CopyGuard defense mechanism.

Supported Models

The framework supports the following LVLMs:

Local Models:

• Qwen2.5-VL (7B, 3B): State-of-the-art multimodal model
  • Install: pip install qwen-vl-utils
• DeepSeek-VL (7B): High-performance vision-language model
  • Install: git clone https://github.com/deepseek-ai/DeepSeek-VL.git && cd DeepSeek-VL && pip install -e .
• LLaVA (1.5-7B, 1.5-13B, Next): Popular open-source VLM
  • Install: git clone https://github.com/haotian-liu/LLaVA.git && export PYTHONPATH=/path/to/LLaVA:$PYTHONPATH
• GLM-4V (9B): Bilingual vision-language model
• InstructBLIP (Vicuna-7B): Instruction-tuned BLIP
• Janus-Pro (7B): Unified vision understanding and generation model

API-Based Models:

• GPT-4o, GPT-4o-mini: OpenAI's multimodal models
• Claude 3 (Opus, Sonnet, Haiku): Anthropic's vision models
• Gemini Pro: Google's multimodal model

Step 1: Generate Baseline Responses (Without Defense)

Generate baseline model responses without any defense mechanism:

python -m models.generate_responses \
    --model-type qwen \
    --model-path /path/to/qwen2.5-vl-7b \
    --dataset dataset/book_copyright.json \
    --output results/book_baseline.json \
    --image-mode 0 \
    --notice-mode 0

Step 2: Evaluate Baseline Responses

Evaluate the generated baseline responses:

python -m evaluation.evaluator \
    --input results/book_baseline.json \
    --output results/book_baseline_eval.json \
    --csv results/book_baseline_metrics.csv

Step 3: Generate Responses With CopyGuard Defense

Generate model responses with the CopyGuard defense mechanism enabled:

python -m models.generate_responses_with_defense \
    --model-type qwen \
    --model-path /path/to/qwen2.5-vl-7b \
    --dataset dataset/book_copyright.json \
    --output results/book_defense.json \
    --image-mode 0 \
    --notice-mode 0

Parameter Explanation:

The benchmark evaluates 5 forms of copyright notices by combining --image-mode and --notice-mode parameters:

Copyright Notice Form	--image-mode	--notice-mode	Description
1. No notice (baseline)	0	0	Plain text image with no copyright notice
2. Generic text notice	0	1	Plain text image with "All rights reserved" in query
3. Original text notice	0	2	Plain text image with original copyright text in query
4. Generic image notice	1	0	Image with "All rights reserved" embedded, no text notice
5. Original image notice	2	0	Image with original copyright text embedded, no text notice

• --image-mode: Specifies the image presentation mode. Options: 0 = plain text image (no copyright notice), 1 = image with generic copyright notice ("All rights reserved"), 2 = image with original copyright notice. This parameter evaluates how the modality of copyright notices (embedded in image vs. presented as text) affects model compliance.
• --notice-mode: Specifies the copyright notice format in the query. Options: 0 = no notice, 1 = generic notice ("All rights reserved"), 2 = original notice (content-specific copyright text). This parameter evaluates how different types of copyright notices impact model behavior, as discussed in the paper.

Step 4: Evaluate Defense-Enabled Responses

Evaluate the responses generated with CopyGuard defense:

python -m evaluation.evaluator \
    --input results/book_defense.json \
    --output results/book_defense_eval.json \
    --csv results/book_defense_metrics.csv

📚 Advanced Usage

Using Different Models

Local Models

Qwen2.5-VL:

python -m models.generate_responses \
    --model-type qwen \
    --model-path /path/to/Qwen2.5-VL-7B-Instruct \
    --dataset dataset/book_copyright.json \
    --output results/qwen_results.json

DeepSeek-VL:

python -m models.generate_responses \
    --model-type deepseek \
    --model-path /path/to/deepseek-vl-7b-chat \
    --dataset dataset/book_copyright.json \
    --output results/deepseek_results.json

LLaVA:

python -m models.generate_responses \
    --model-type llava \
    --model-path /path/to/llava-v1.5-7b \
    --llava-repo-path /path/to/LLaVA \
    --dataset dataset/book_copyright.json \
    --output results/llava_results.json

GLM-4V:

python -m models.generate_responses \
    --model-type glm \
    --model-path /path/to/glm-4v-9b \
    --dataset dataset/book_copyright.json \
    --output results/glm_results.json

API-Based Models

GPT-4o:

export OPENAI_API_KEY="your-api-key"
python -m models.generate_responses \
    --model-type gpt-4o \
    --model-path gpt-4o \
    --dataset dataset/book_copyright.json \
    --output results/gpt4o_results.json

Claude:

export ANTHROPIC_API_KEY="your-api-key"
python -m models.generate_responses \
    --model-type claude \
    --model-path claude-3-7-sonnet-20250219 \
    --api-base https://api.anthropic.com/v1 \
    --dataset dataset/book_copyright.json \
    --output results/claude_results.json

CopyGuard Defense Options

The CopyGuard defense framework can be customized with various options:

python -m models.generate_responses_with_defense \
    --model-type qwen \
    --model-path /path/to/qwen2.5-vl-7b \
    --dataset dataset/book_copyright.json \
    --output results/book_defense.json \
    --enable-ocr \              # Enable OCR-based copyright detection
    --enable-verifier \         # Enable copyright status verification
    --enable-risk-analyzer \    # Enable query risk analysis
    --enable-reminder \         # Enable copyright reminders
    --block-risky              # Block risky queries (vs. warning only)

Defense Components:

• Notice Identifier: Uses PaddleOCR to detect copyright notices in images
• Status Verifier: Uses Serper API + DeepSeek-R1 to verify copyright status
• Risk Analyzer: Analyzes queries for potential copyright infringement
• Status Reminder: Provides copyright guidance to the model

Batch Processing

Process all 5 copyright notice configurations:

#!/bin/bash
MODEL_TYPE="qwen"
MODEL_PATH="/path/to/qwen2.5-vl-7b"
DATASET="dataset/book_copyright.json"

for image_mode in 0 1 2; do
    for notice_mode in 0 1 2; do
        if [ $image_mode -eq 0 ] || [ $notice_mode -eq 0 ]; then
            echo "Processing image_mode=$image_mode, notice_mode=$notice_mode"
            python -m models.generate_responses \
                --model-type $MODEL_TYPE \
                --model-path $MODEL_PATH \
                --dataset $DATASET \
                --output results/book_${image_mode}_${notice_mode}.json \
                --image-mode $image_mode \
                --notice-mode $notice_mode
        fi
    done
done

Dataset Structure

The dataset JSON should follow this structure:

[
    {
        "txt_file": "dataset/book_copyright/1Q84.txt",
        "img_file_0": "dataset/book_copyright_images/0/1Q84_sample_1.png",
        "img_file_1": "dataset/book_copyright_images/1/1Q84_sample_1.png",
        "img_file_2": "dataset/book_copyright_images/2/1Q84_sample_1.png",
        "text": "Full text content from the book...",
        "copyright_text": "Copyright © 2009 by Haruki Murakami. All rights reserved."
    }
]

Each entry contains:

• txt_file: Path to the text file
• img_file_0/1/2: Paths to images with different copyright notice modes
• text: The actual text content
• copyright_text: Original copyright notice text

📄 Citation

If you find this work useful for your research, please cite our paper:

@inproceedings{xu2026bridging,
  title={Bridging the Copyright Gap: Do Large Vision-Language Models Recognize and Respect Copyrighted Content?},
  author={Xu, Naen and Zhang, Jinghuai and Li, Changjiang and An, Hengyu and Zhou, Chunyi and Wang, Jun and Xu, Boyu and Li, Yuyuan and Du, Tianyu and Ji, Shouling},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  year={2026}
}