Double Visual Defense

RESOURCE

Abstract

This is the official code for the paper "Double Visual Defense: Adversarial Pre-training and Instruction Tuning for Improving Vision-Language Model Robustness". This code can be used to produce vision language models (VLMs), like LLaVA, with enhanced robustness to adversarial attacks (e.g. jailbreaks).
Developers:
Bartoldson, Brian [1] Wang, Zeyu [1] Kailkhura, Bhavya [1]
  1. Lawrence Livermore National Laboratory (LLNL), Livermore, CA (United States)
Release Date:
2024-12-12
Project Type:
Open Source, Publicly Available Repository
Software Type:
Scientific
Version:
1
Licenses:
MIT License
Sponsoring Org.:
Code ID:
154189
Site Accession Number:
LLNL-CODE-2002980
Research Org.:
Lawrence Livermore National Laboratory (LLNL), Livermore, CA (United States)
Country of Origin:
United States

RESOURCE

Citation Formats

Bartoldson, Brian, Wang, Zeyu, and Kailkhura, Bhavya. Double Visual Defense. Computer Software. https://github.com/zw615/Double_Visual_Defense. USDOE National Nuclear Security Administration (NNSA). 12 Dec. 2024. Web. doi:10.11578/dc.20250416.3.
Bartoldson, Brian, Wang, Zeyu, & Kailkhura, Bhavya. (2024, December 12). Double Visual Defense. [Computer software]. https://github.com/zw615/Double_Visual_Defense. https://doi.org/10.11578/dc.20250416.3.
Bartoldson, Brian, Wang, Zeyu, and Kailkhura, Bhavya. "Double Visual Defense." Computer software. December 12, 2024. https://github.com/zw615/Double_Visual_Defense. https://doi.org/10.11578/dc.20250416.3.
@misc{ doecode_154189,
title = {Double Visual Defense},
author = {Bartoldson, Brian and Wang, Zeyu and Kailkhura, Bhavya},
abstractNote = {This is the official code for the paper "Double Visual Defense: Adversarial Pre-training and Instruction Tuning for Improving Vision-Language Model Robustness". This code can be used to produce vision language models (VLMs), like LLaVA, with enhanced robustness to adversarial attacks (e.g. jailbreaks).},
doi = {10.11578/dc.20250416.3},
url = {https://doi.org/10.11578/dc.20250416.3},
howpublished = {[Computer Software] \url{https://doi.org/10.11578/dc.20250416.3}},
year = {2024},
month = {dec}
}