Compression-Aware Counterfactual Defense

Overview

Implementation of compression-aware counterfactual defense against prompt compression attacks, as described by Liu et al.

Project Structure

• compression_aware_defense.py - Core defense implementation
• token_level_attack.py - HardCom token-level attack
• qa_eval.py - Question answering task evaluation (SQuAD)
• integrated_evaluation.py - Complete defense evaluation framework

Setup

pip install -r requirements.txt
export OPENAI_API_KEY="your-key-here"

Running Experiments

Product Recommendation Attack

python product_rec_eval.py

QA Attack (SQuAD)

python qa_eval.py

Defense Evaluation

python integrated_evaluation.py

Results

• integrated_evaluation_results.json - Defense performance metrics
• squad_qa_attack_results.json - QA attack results across compression rates

Requirements

• Python 3.10+
• OpenAI API key
• See requirements.txt for dependencies

Citation

@misc{liu2025compressionattackexploitingpromptcompression,
      title={CompressionAttack: Exploiting Prompt Compression as a New Attack Surface in LLM-Powered Agents}, 
      author={Zesen Liu and Zhixiang Zhang and Yuchong Xie and Dongdong She},
      year={2025},
      eprint={2510.22963},
      archivePrefix={arXiv},
      primaryClass={cs.CR},
      url={https://arxiv.org/abs/2510.22963}, 
}