ahnpersie/llama3.1-8b-lora-coco-deceptive-clip

LLaMA-3.1-8B-LoRA-COCO-Deceptive-CLIP Model Card

🏆 This work is accepted to ACL 2025 (Main Conference).

<p align="left"> <img src="./main_result.png" alt="main result" width="60%" height="60%"> <em>Figure: Attack success rate (ASR) and caption diversity of our model on the COCO dataset, illustrating its ability to generate deceptive captions that successfully fool CLIP.</em> </p>

Model Description

• —Repository: Code
• —Paper: Can LLMs Deceive CLIP? Benchmarking Adversarial Compositionality of Pre-trained Multimodal Representation via Text Updates
• —Point of Contact: Jaewoo Ahn, Heeseung Yun

Dataset

This model was fine-tuned on the [COCO-Deceptive-CLIP-LLaMA-3.1-8B Training Dataset](https://huggingface.co/datasets/ahnpersie/coco-deceptive-clip-llama3.1-8b), which provides structured instruction–response pairs for generating deceptive captions that mislead CLIP.

Model Details

• —Model: LLaMA-3.1-8B-LoRA-COCO-Deceptive-CLIP is a deceptive caption generator built on LLaMA-3.1-8B, fine-tuned using LoRA (i.e., self-training, or more specifically, rejection sampling fine-tuning (RFT)) to deceive CLIP on the COCO dataset. It achieves an attack success rate (ASR) of 42.1%.
• —Architecture: This model is based on LLaMA-3.1-8B and utilizes PEFT v0.12.0 for efficient fine-tuning.

How to Use

See our GitHub repository for full usage instructions and scripts.

Citation

Please cite our work if you find the resources in this repository useful:

@inproceedings{ahn2025mac,
      title={Can LLMs Deceive CLIP? Benchmarking Adversarial Compositionality of Pre-trained Multimodal Representation via Text Updates},
      author={Jaewoo Ahn and Heeseung Yun and Dayoon Ko and Gunhee Kim},
      booktitle={ACL},
      year=2025
}