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

COCO-Deceptive-CLIP-LLaMA-3.1-8B Training Dataset

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

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.

Dataset Details

This dataset provides instruction–response pairs formatted as short two-turn conversations:

• The user message contains:

  • A given image caption.
  • A set of task instructions defining the deceptive caption generation rules.
  • The requirement to output a Generated Caption: that contradicts the original caption while remaining semantically close enough to fool CLIP.

• The assistant message contains:

  • A single line that begins with

    Generated Caption: ...
    

    which serves as the synthesized “deceptive (or adversarial)" caption.

Each dataset instance follows the structure:

[
  {
    "content": "<deceptive caption generation instructions + given caption>",
    "role": "user"
  },
  {
    "content": "Generated Caption: <model-generated deceptive caption>",
    "role": "assistant"
  }
]

This conversational schema is optimized for fine-tuning instruction-following models to produce deceptive captions that increase CLIP similarity while contradicting the ground-truth semantics with minimal word-level edits.

---

Relation to the Fine-Tuned Model

This dataset was used to fine-tune 👉 ahnpersie/llama3.1-8b-lora-coco-deceptive-clip, a LoRA-adapted version of LLaMA-3.1-8B that learns to generate deceptive captions capable of misleading CLIP.

The released model demonstrates:

• effective adversarial caption generation,
• a strong attack success rate (ASR) & attack diversity (H) on COCO,
• and improved compositional deception behavior originating from this dataset’s structured supervision.

---

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
}