⚔️ Adversarial Attacks against Closed-Source MLLMs via Feature Optimal Alignment

FOA-Attack is proposed to enhance adversarial transferability in multimodal large language models by optimizing both global and local feature alignments using cosine similarity and optimal transport.

💥 News

• [2025-09-19] Our paper has been accepted to NeurIPS 2025! 🎉
• [2025-05-29] We release the FOA-Attack code! 🚀

💻 Requirements

Dependencies: To install requirements:

pip install -r requirements.txt

🛰️ Quick Start

python generate_adversarial_samples_foa_attack.py
python blackbox_text_generation.py -m blackbox.model_name=gpt4o,claude,gemini
python gpt_evaluate.py -m blackbox.model_name=gpt4o,claude,gemini
python keyword_matching_gpt.py -m blackbox.model_name=gpt4o,claude,gemini

⚙️ Start (automatic selection of cluster centers)

python FOAttack.py

💖 Acknowledgements

This project is built on M-Attack. We sincerely thank them for their outstanding work.