R.A.C.E.: Robust Adversarial Concept Erasure for Secure Text-to-Image Diffusion Model (ECCV'24 Oral)

In the evolving landscape of text-to-image (T2I) diffusion models, the remarkable capability to generate high-quality images from textual descriptions faces challenges with the potential misuse of reproducing sensitive content. To address this critical issue, we introduce Robust Adversarial Concept Erase (RACE), a novel approach designed to mitigate these risks by enhancing the robustness of concept erasure method for T2I models. RACE utilizes a sophisticated adversarial training framework to identify and mitigate adversarial text embeddings, significantly reducing the Attack Success Rate (ASR). Impressively, RACE achieves a 30% reduction in ASR for the "nudity" concept against the leading white-box attack method. Our extensive evaluations demonstrate RACE's effectiveness in defending against both white-box and black-box attacks, marking a significant advancement in protecting T2I diffusion models from generating inappropriate or misleading imagery. This work underlines the essential need for proactive defense measures in adapting to the rapidly advancing field of adversarial challenges.

Authors

Changhoon Kim
Kyle Min
Yezhou Yang

News

Arxiv Link: https://arxiv.org/abs/2405.16341
(2024.7.1) R.A.C.E. is accepted at ECCV.

Motivation

Comparative demonstration of concept erasure, red teaming, and robust erasure within T2I diffusion models. The ESD method removes targeted concepts from the original SD outputs, yet these concepts can be reconstructed using UnlearnDiff. Our proposed R.A.C.E. method showcases enhanced robustness against such red teaming reconstruction efforts.

Training Configuration

Our framework builds upon the Erasing Concepts from Diffusion Models project. For initial setup and training configuration, kindly refer to the provided link.

Guide for Adversarial Training

First, run the code from ESD, and then run this code.
The main code for adversarial training can be found in /train-scripts/train-esd.py.

To engage in adversarial training following the initial model setup with ESD, please integrate the following parameters:

• iteration - specifies the number of iterations for
• adv_train - a flag to activate adversarial training mode.
• exp_name - designates the name of the experiment.
• pretrained_esd- specifies the path to the pre-trained ESD model weights.
• epsilon - defines the magnitude of the epsilon ball.
• pgd_num_step - sets the number of steps for the adversarial attack process.
• lasso - a flag for the regularization method discussed in the Discussion section.

Please note the necessity to adjust the default values in the argument parser.