[ECCV 2024]
UnlearnDiffAtk:
To Generate or Not?
Safety-Driven Unlearned Diffusion Models
Are Still Easy To Generate Unsafe Images
... For Now

Project Website | Arxiv Preprint | Unlearned DM Benchmark | Demo

Welcome to the official implementation of UnlearnDiffAtk, which capitalizes on the intrinsic classification abilities of Diffusion Models (DMs) to simplify the creation of adversarial prompts, thereby eliminating the need for auxiliary classification or diffusion models.Through extensive benchmarking, we evaluate the robustness of five widely-used safety-driven unlearned DMs (i.e., DMs after unlearning undesirable concepts, styles, or objects) across a variety of tasks.

Abstract

The recent advances in diffusion models (DMs) have revolutionized the generation of complex and diverse images. However, these models also introduce potential safety hazards, such as the produc- tion of harmful content and infringement of data copyrights. Although there have been efforts to create safety-driven unlearning methods to counteract these challenges, doubts remain about their capabilities. To bridge this uncertainty, we propose an evaluation framework built upon adversarial attacks (also referred to as adversarial prompts), in order to discern the trustworthiness of these safety-driven unlearned DMs. Specifically, our research explores the (worst-case) robustness of unlearned DMs in eradicating unwanted concepts, styles, and objects, assessed by the generation of adversarial prompts. We develop a novel adversarial learning approach called UnlearnDiff that leverages the inherent classification capabilities of DMs to streamline the generation of adversarial prompts, making it as simple for DMs as it is for image classification attacks. This technique streamlines the creation of adversarial prompts, making the process as intuitive for generative modeling as it is for image classification assaults. Through comprehensive benchmarking, we assess the unlearning robustness of five prevalent unlearned DMs across multiple tasks. Our results underscore the effectiveness and efficiency of UnlearnDiff when compared to state-of-the-art adversarial prompting methods

UnlearnDiffAtk: Unlearned Diffusion Model Benchmark

DM Unlearning Tasks:

• NSFW: Nudity
• Style: Van Gogh
• Object: Church, Tench, Parachute, Garbage Truck

Evaluation Metrics:

• Pre-Attack Success Rate (Pre-ASR): lower is better;
• Post-Attack Success Rate (Post-ASR): lower is better;
• Fréchet inception distance(FID): evaluate distributional quality of image generations, lower is better;
• CLIP Score: measure contextual alignment with prompt descriptions, higher is better.

Unlearned Concept: Nudity

Unlearned Methods	Pre-ASR (%)	Post-ASR (%)	FID	CLIP-Score
EraseDiff (ED)	0.00	2.11	233	0.18
ScissorHands (SH)	0.00	7.04	128.53	0.235
Saliency Unlearning (SalUn)	1.41	11.27	33.62	0.287
Adversarial Unlearning (AdvUnlearn)	7.75	21.13	19.34	0.290
Erased Stable Diffusion (ESD)	20.42	76.05	18.18	0.302
Unified Concept Editing (UCE)	21.83	79.58	17.10	0.309
Forget-Me-Not (FMN)	88.03	97.89	16.86	0.308
concept-SemiPermeable Membrane (SPM)	54.93	91.55	17.48	0.310

We will evaluate your model on UnlearnDiffAtk Benchmark!

Open a github issue or email us at zhan1853@msu.edu!

FULL UnlearnDiffAtk Unlearned DM benchmark can be found in Huggingface!

Code Structure

configs: contains the default parameter for each methods

prompts: contains the prompts we selected for each experiments

src: contains the source code for the proposed methods

• attackers: contains different attack methods (different discrete optimization methods)
• tasks: contains different type of attacks (auxiliary model-based attacks P4D, and ours UnlearnDiff)
• execs: contains the main execution files to run experiments
• loggers: contains the logger codes for the experiments

Prepare

Environment Setup

conda env create -n ldm --file environments/x86_64.yaml

Unlearned model preparation

We provide different unlearned models (ESD and FMN), and you can download them from [Object , Others]. We also provide an Artist classifier for evaluating the style task. You can download it from here.

Generate dataset (for unlearning robustness evaluation)

python src/execs/generate_dataset.py --prompts_path prompts/nudity.csv --concept i2p_nude --save_path files/dataset

Code Implementation

In this section, we provide the instructions to reproduce the results on nudity (ESD) in our paper. You can change the config file path to reproduce the results on other concepts or unlearned models.

No attack

python src/execs/attack.py --config-file configs/nudity/no_attack_esd_nudity_classifier.json --attacker.attack_idx $i --logger.name attack_idx_$i

where i is from [0,142)

UnlearnDiff attack

python src/execs/attack.py --config-file configs/nudity/text_grad_esd_nudity_classifier.json --attacker.attack_idx $i --logger.name attack_idx_$i

where i is from [0,142)

Evaluation

For nudity/violence/illegal/objects:

python scripts/analysis/check_asr.py --root-no-attack $path_to_no_attack_results --root $path_to_${P4D|UnlearnDiff}_results

For style:

python scripts/analysis/style_analysis.py --root $path_to_${P4D|UnlearnDiff}_results --top_k {1|3}

Citation

@article{zhang2023generate,
  title={To Generate or Not? Safety-Driven Unlearned Diffusion Models Are Still Easy To Generate Unsafe Images... For Now},
  author={Zhang, Yimeng and Jia, Jinghan and Chen, Xin and Chen, Aochuan and Zhang, Yihua and Liu, Jiancheng and Ding, Ke and Liu, Sijia},
  journal={arXiv preprint arXiv:2310.11868},
  year={2023}
}

Related Works - Machine Unlearning

• Adversarial Unlearning (AdvUnlearn)
• Sparse Unlearning (l1-sparse)
• Saliency Unlearning (SalUn)