DPM: Copyright Infringement Detection in Text-to-Image Diffusion Models via Differential Privacy

---

This repository contains the official implementation of D-Plus-Minus (DPM), a principled post-hoc framework for detecting copyright infringement in text-to-image diffusion models, grounded in differential privacy.

Environment Setup

We recommend using Anaconda to set up a clean environment:

cd DPM-copyright-infringement-detection

# Create conda environment
conda create -n dpm python=3.12 -y
conda activate dpm

# Install dependencies
pip install -r requirements.txt

Configuration

Edit the config YAML in ./configs/ folder before running:

Config Field	Description
images_class	Must be one of: human_face, architecture, arts_painting
images_name	Name of the subfolder (concept) under ./data/CIDD-example/class/<class>/
output_folder	Directory where the model checkpoints, generated images, detection results, and logs will be saved.

⚠️ Do not modify other config values, especially those related to normalization and scoring, unless you know what you're doing.

Changing them may lead to incorrect normalization in conditional sensitivity scoring and compromise the validity of the DPM confidence scores.

Running DPM Detection

Choose one of the pre-defined models and run:

export CONFIG_NAME="sd_1_4"
python main.py --config-name $CONFIG_NAME

Supported Model Configs:

CONFIG_NAME	Model Description
sd_1_4	Stable Diffusion v1.4 (full fine-tuning)
lora_sdxl	Stable Diffusion XL (1.0) + LoRA
lora_sana	SANA (0.6B) + LoRA
lora_flux	FLUX.1 + LoRA

All model-specific settings are defined in the corresponding YAML files in ./configs.