This repository contains the code and experiments for the paper:

"Inverting Black-Box Face Recognition Systems via Zero-Order Optimization in Eigenface Space"

🔧 How to Use

📁 Datasets

• LFW
  Download lfw.tgz and place it into the data/ directory.

• AgeDB-30
  No manual setup required — the script will automatically download and extract it.

• CFP-FP
  Download cfp-dataset.zip and place it into the data/ directory.

---

🧠 Models

• ArcFace Backbone
  Download ms1mv3_arcface_r100_fp16_backbone.pth and place it into the models/ directory.

• Eigenfaces & Average Face (for reconstruction)
  Download the following files and place them into the models/ directory:

  • eigenvectors_denormalized_Square_ffhq.pt
  • mtcnn_crops_ffhq_average_vector_square.pt

🚀 Running the Full Pipeline

1. Step 1: Prepare datasets and evaluate on original images

   bash scripts/download_preprocess_eval_on_original_sets.sh

   • Unpacks datasets
   • Aligns all faces (MTCNN / RetinaFace)
   • Runs baselines with ArcFace and FaceNet on original (real) images

2. Step 2: Reconstruct positive 1st faces in positive pairs with our method

   bash scripts/reconstruct_faces.sh

   • It takes ~30 hours on 4×A100 GPUs to run
   • Runs our zero-order optimization algorithm
   • Restores the first face in all positive verification pairs
   • Supports ArcFace-based and FaceNet-based reconstructions

3. Step 3: Evaluate reconstructed faces

   bash scripts/run_eval_on_reconstructed.sh

   • Aligns reconstructed images (MTCNN / RetinaFace)
   • Runs ArcFace and FaceNet verification on restored faces

---

📝 Results are saved into:

runs/results.txt

⚙️ Optional: Generate Your Own PCA Basis

If you prefer to generate your own PCA eigenfaces instead of using the precomputed ones, run the following script:

python algorithms/generate_pca_basis.py

You need to specify the path to a folder with face images. We used the FFHQ thumbnails dataset (thumbnails128x128 version).

The script will save:

• A mean face vector mtcnn_crops_ffhq_average_vector_square.pt
• A PCA projection matrix eigenvectors_denormalized_Square_ffhq.pt (eigenfaces)

Make sure to copy the generated files to the models/ folder so that reconstruction scripts can find them.