t-maho / fbi_fingerprinting

FBI

Code to Fingerprint any model with Benign Inputs

Due to the lack of space, only Top-1 data is released in the GitHub.

Some results of our paper need images. We provide a small set of 1000 images. These images are the ones with the highest entroyp score.

Run

First, install requirements

pip install -r requirements

Known

Detection

python known_detection.py --input ./data/predictions --score mean_case --family pure --max_drop -0.15

family argument can be choosen between: pure, variation, singleton

Identification

python known_identification.py --input ./data/predictions --score mean_case --family pure --max_drop -0.15

family argument can be choosen between: pure, variation, singleton

Unknown

Detection

python unknown_detection.py --input ./data/predictions/ --output_dir ./results/ --delegate close --family pure --n_images 10 50 100 150 200

Identification Pure

python unknown_detection.py --input ./data/predictions/ --output_dir ./results/ --delegate close --family pure --n_images 10 50 100 150 200

Identification Variation (with compound)

python unknown_identification_variation.py --input ./data/predictions/ --output_dir ./results/ --delegate middle close --n_images 50 100 200 --sort_images entropy --n_gen 20