Ring-A-Bell! How Reliable are Concept Removal Methods For Diffusion Models?

Authors: Yu-Lin Tsai, Chia-Yi Hsu, Chulin Xie, Chih-Hsun Lin, Jia You Chen, Bo Li, Pin-Yu Chen, Chia-Mu Yu, Chun-Ying Huang

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our extensive experiments evaluate a wide range of models, ranging from popular online services to state-of-the-art concept removal methods, and reveal that problematic prompts generated by Ring-A-Bell can increase the success rate for most concept removal methods in generating inappropriate images by more than 30%.
Researcher Affiliation Collaboration Chia-Yi Hsu , Yu-Lin Tsai National Yang Ming Chiao Tung University {chiayihsu8315,uriah1001}@gmail.com Chulin Xie University of Illinois at Urbana Champaign chulinx2@illinois.edu Chih-Hsun Lin, Jia-You Chen National Yang Ming Chiao Tung University {pkevawin334, justin041510}@gmail.com Bo Li University of Illinois at Urbana Champaign University of Chicago lbo@illnois.edu, bol@uchicago.edu Pin-Yu Chen IBM Research pin-yu.chen@ibm.com Chia-Mu Yu, Chun-Ying Huang National Yang Ming Chiao Tung University chiamuyu@gmail.com, chuang@cs.nctu.edu.tw
Pseudocode No The paper describes its methods but does not include explicit pseudocode or algorithm blocks.
Open Source Code Yes Our codes are available at https://github.com/chiayi-hsu/Ring-A-Bell.
Open Datasets Yes Dataset. We evaluate the performance of Ring-A-Bell on the I2P dataset (Schramowski et al., 2023), an established dataset of problematic prompts, on the concepts of nudity and violence.
Dataset Splits No The paper describes selecting subsets of prompts from the I2P dataset for evaluation but does not specify train, validation, or test splits for its own experimental methodology.
Hardware Specification No The paper does not provide specific details about the hardware used to run the experiments.
Software Dependencies No The paper mentions using CLIP with VIT-L/14 but does not specify version numbers for other key software dependencies or libraries.
Experiment Setup Yes To run the GA, we use 200 random initial prompts with 3000 generations and set the mutation rate and crossover rate to 0.25 and 0.5, respectively. Furthermore, there are hyper-parameters: K (the length of the prompts), η (the weight of the empirical concept), and N (the number of prompt pairs). In Section 4.3, we will show how K, η, N as well as the choice of optimizer affect the attack results.