/HufuNet

Primary LanguagePython

[HufuNet: Embedding The Left Piece as Watermark and Keeping The Right Piece for Ownership Verification in Deep Neural Networks]

A PyTorch introduction of HufuNet.

Introduction

Given a watermark as HufuNet, the following is the workflow of HufuNet:

alt text

Usage

Generate HufuNet:

python train.py --model==FLenet --checkpoint==FLenet --epochs=100  ==>  FLenet.t7 (92.60%)

Embedded HufuNet into VGG:

python train.py --model=VGG --checkpoint=VGG_init --epochs=1  ==>  VGG_init.t7 (39.59%)
python vgg_grad_false.py  ==>  VGG_done_3*3.t7(Final Poisoned Model 77.57%)

Simulate Attacking Watermarked-VGG Model:

python vgg_grad_true.py  ==>  VGG_done_finetune100.t7 (Finetune 75.15%)
python prune.py --prune_rate=60  ==>  VGG_done_3*3_prune60.t7 (Prune 71.65%)
python simulate_reorder_attack.py  ==>  VGG_ReOrder.t7 (Reorder 71.65%)
python simulate_EnlargeC_attack.py  ==>  VGG_EnlargeC.t7 (Final Victim Model 71.65%)

Restore VGG to Origin:

python simulate_restore_scale.py  ==>  VGG_RestoreCoff.t7 (Restore Number Scale)
python simulate_restore_order.py  ==>  VGG_RestoreOrder.t7 (Restore Filters Position)

Extract HufuNet from VGG and verify it on FASHION-MNIST:

python structureDiff_and_convertEmbed.py --model1=FLenet --model2=VGG_RestoreOrder

Draw Cos-Similarity Distribution of VGG:

python draw_parameters_distribution.py