Inverse Embedding Attack

Realistic Inverse Embedding Attack Solution

Main Idea

Instead of training attackers directly on black-box models (unrealistic), we:

1. Train attackers on similar embedding models
2. Use transfer learning to attack black-box models
3. No need for original black-box model data

Main Flow

Dataset → Multiple Embedding Models → Train Attackers → Test on Black-box

Project Structure

Inverse_Embedding_Attack/
├── src/
│   ├── data_processing/     # Dataset processing and embedding creation
│   ├── attackers/          # Attacker models
│   ├── evaluation/         # Result evaluation
│   └── utils/             # Utilities and helpers
├── data/                  # Datasets and embeddings
├── models/               # Trained models
├── experiments/          # Configs and results
└── configs/             # Configuration files

Supported Models

• Embedding Models: all-mpnet-base-v2, stsb-roberta-base, all-MiniLM-L6-v2, paraphrase-MiniLM-L6-v2
• Attacker Models: GPT-2, OPT, T5 (from GEIA)
• Datasets: SST-2, PersonaChat, ABCD

Usage

# 1. Prepare data and embeddings
python src/data_processing/prepare_embeddings.py

# 2. Train attackers
python src/attackers/train_attackers.py

# 3. Test on black-box
python src/evaluation/test_blackbox.py