SOPHON: Non-Fine-Tunable Learning to Restrain Task Transferability For Pre-trained Models

https://arxiv.org/abs/2404.12699

Jiangyi Deng (1), Shengyuan Pang (1), Yanjiao Chen (1), Liangming Xia (1), Yijie Bai (1), Haiqin Weng (2), Wenyuan Xu (1) ((1) Zhejiang University, (2) Ant Group)

Accepted by IEEE Symposium on Security and Privacy 2024

Table of Contents

• Introduction

• Preperation

• Usage

Introduction

This is the implementation of our paper: SOPHON: Non-Fine-Tunable Learning to Restrain Task Transferability For Pre-trained Models

Preperation

You can build the required environment by running:

conda env create -f environment.yml

Put pretrained models in ./classification/pretrained and ./generation/pretrained

Put datasets in ../datasets

Usage

The whole project is devided into two parts:

• classification : codes for reproducing our classification-related experiments
• generation : codes for reproducing our generation-related experiments

Classification task

Workspace is ./classification, thus

cd classification

Train Sophoned model

For inverse cross-entropy sophon, run:

python inverse_loss.py --alpha 3 --beta 5 --dataset CIFAR10 --arch res50

The output ckpt will be saved to results/inverse_loss/[args.arch]_[args.dataset]/[current_time]/

For kl divergence from uniform distribution sophon, run:

python kl_uniform_loss.py.py --alpha 1 --beta 1 --nl 5 --dataset CIFAR10 --arch res50

The choices of args.dataset are [CIFAR10, CINIC, SVHN, STL, MNIST]

The choices of args.arch are [caformer, res50, res34, res18, vgg]

The output ckpt will be saved to results/kl_loss/[args.arch]_[args.dataset]/[current_time]/

Test finetune

For test a target ckpt's finetune outcome directly:

# for finetuned ckpt
python finetune_test.py --start sophon --path path_to_ckpt

# for normal pretrained
python finetune_test.py --start normal

# for train from scratch
python finetune_test.py --start scratch

Generation task

Workspace is ./generation, thus:

cd generation

Train Sophoned model

For mean squred loss sophon, run:

python ./mean_squared_loss.py --alpha 1.0 --beta 5.0 --bs 100 --fast_batches 50 --ml 1 --nl 1

The output ckpts will be saved to: ./res/mean_squared_loss_celeba/[current time]/

For denial service loss sophon, run:

python denial_service_loss.py --alpha 0.05 --beta 2 --nl 10 --total 200

The output ckpts will be saved to: ./res/denial_service_loss_celeba/[current time]/

Test finetune

For test a target ckpt's finetune outcome directly:

For any sophoned or other processed ckpt

run:

# for mean squared loss test
python mean_squared_loss.py --finetune path_to_ckpt 

# for denial service loss test
python denial_service_loss.py --finetune path_to_ckpt  

For two baselines: normal pretrained ckpt or train from scratch

run:

# for normal pretrained baseline
python mean_squared_loss.py --pretest scratch

# for train from scratch baseline
python mean_squared_loss.py --pretest pretrained