Jailbreak in pieces: Compositional Adversarial Attacks on Multi-Modal Language Models

Authors: Erfan Shayegani, Yue Dong, Nael Abu-Ghazaleh

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our attacks achieve a high success rate for two different VLMs we evaluated, highlighting the risk of cross-modality alignment vulnerabilities, and the need for new alignment approaches for multi-modal models.
Researcher Affiliation Academia Erfan Shayegani, Yue Dong & Nael Abu-Ghazaleh Department of Computer Science University of California, Riverside {sshay004,yued,naelag}@ucr.edu
Pseudocode Yes Algorithm 1: Adversarial Image Generator via Embedding Space Matching
Open Source Code No The paper states: 'We plan to release our dataset with 4 types of malicious triggers and prompts.' (Footnote 1), but does not explicitly state the release of the source code for the methodology or provide a link.
Open Datasets Yes Zou et al. (2023) and Bailey et al. (2023) utilize Adv Bench, which consists of 521 lines of harmful behaviors and 575 lines for harmful strings.
Dataset Splits No The paper describes testing repetitions ('repeating each experiment 25 times') and the scenarios used, but does not provide specific train/validation/test dataset splits or reference predefined splits for reproducibility.
Hardware Specification Yes typically within 10 to 15 minutes when utilizing a Google Colab T4 GPU
Software Dependencies No The paper mentions using specific CLIP models and the ADAM optimizer, but does not provide specific software dependencies with version numbers (e.g., Python 3.8, PyTorch 1.9, etc.).
Experiment Setup Yes Out of our experiments, we empirically found that a distance around 0.3 or lower often indicates a powerful adversarial sample that will be as effective as the target malicious trigger when presented to the VLM.