Unveiling the Attribute Misbinding Threat in Identity-Preserving Models

Table of Contents

• Unveiling the Attribute Misbinding Threat in Identity-Preserving Models
  • Framework Overview
  • Attack Demonstration
  • Identity-Preserving Models
  • Text Filter
  • Safety Evaluation
  • Quick Start
    • Installation
    • Sensitive Term Expansion
    • Attribute Misbinding Attack
    • Attribute Binding Safety Score

Framework Overview

The proposed framework for generating Misbinding Prompt evaluation set and evaluating the safety of identity-preserving models.
The framework consists of four stages:

1. Sensitive Term Expansion
   Methodically broaden the vocabulary of sensitive terms.
2. Attribute Misbinding Attack
   Programmatically create prompts via predefined strategies.
3. Diffusion Generation
   Use prompts and identity reference images for synthesis.
4. Attribute Binding Safety Score Calculation
   An MLLM assesses the output to calculate the final score.

---

Attack Demonstration

Demonstration of the proposed Attribute Misbinding Attack against five leading identity-preserving models.
To avoid infringing upon the portrait rights of real individuals, all reference face images used in this demonstration are portraits generated by StyleGAN2.

Identity-Preserving Models

The Identity-Preserving Models tested in the paper are as follows:

Model	File name	Paper or technical report
UniPortrait	UniPortrait_gen.py	UniPortrait: A Unified Framework for Identity-Preserving Single- and Multi-Human Image Personalization [ICCV2025]
InfiniteYou	InfiniteYou_gen.py	InfiniteYou: Flexible Photo Recrafting While Preserving Your Identity [ICCV 2025 Highlight]
PuLID	PuLID_gen	PuLID: Pure and Lightning ID Customization via Contrastive Alignment [NeurIPS 2024]
PhotoMaker	PhotoMaker_gen.py	PhotoMaker: Customizing Realistic Human Photos via Stacked ID Embedding [CVPR 2024]
FLUX.1 Kontext	Flux-kontext_gen.py	FLUX.1 Kontext [dev] is a 12 billion parameter rectified flow transformer capable of editing images based on text instructions.

Text Filter

The Text Filter used in the paper is as follows:

Method	File name	GitHub/Hugging Face Pages
NSFW-TC	NSFW_TC.py	Fine-tuned DistilRoBERTa-base for NSFW Classification.
Latent Guard	LatenGuard_fliter.py	Democratic AI Safety: No one can define safety for all - Latent Guard helps you define yours.
Detoxify	detoxify_fliter.py	Toxic Comment Classification with ⚡ Pytorch Lightning and 🤗 Transformers.
LLM Template	qwen3_textfilen.py

Safety Evaluation

The detectors used for Safety Evaluation in the paper are as follows:

Method	File name	GitHub/Hugging Face Pages
Q16	q16_evalution.py	Can Machines Help Us Answering Question 16 in Datasheets, and In Turn Reflecting on Inappropriate Content?
Flux-Fliter	flux-fliter.py	Official inference repo for FLUX.1 models.
MLLM Template	mllm_imgfliter.py

Quick Start

Installation

You can run the following script to configure the necessary environment:

conda create -n AMA python=3.10.4
conda activate AMA
pip install -r requirement.txt

Sensitive Term Expansion

python sensitive_term_expansion.py \
    --input "/path/to/your/source_data.xlsx" \
    --output "/path/to/your/custom_output.xlsx" \
    --model_path "HuggingFace_User/Model_Name" \
    --start_row x \
    --end_row y

Attribute Misbinding Attack

python Attribute_Misbinding_Attack/attribute_misbinding.py \
    --input "/path/to/another/input.xlsx" \
    --output "/path/to/results_500_1000.xlsx" \
    --model "/path/to/your/local_model" \
    --start x \
    --end y

Attribute Binding Safety Score

python Attribute_Misbinding_Attack/abss_calculate.py \
    --input_xlsx "/path/to/another/input.xlsx" \
    --image_dir "/path/to/your/images_directory" \
    --model_path "/path/to/your/local_model" \
    --prompt_col "my_prompts" \
    --image_col "my_images" \
    --start x \
    --end y