DP-Forward

Code for the CCS'23 submission "DP-Forward: Fine-tuning and Inference on Language Models with Differential Privacy in Forward Pass"

Note that this repo is anonymous and only intended for review purpose only.

Introduction

We propose DP-Forward, which perturbs embedding matrices in the forward pass of LMs and ensures local DP for training (and test) data.

Setup Environment

Install required packages

pip install -r requirements.txt

Fine-tune BERT model with DP-Forward

python run_glue.py \
--model_name_or_path bert-base-uncased \
--task_name sst2
--do_train
--do_eval
--max_seq_length 64
--per_device_train_batch_size 32
--per_device_eval_batch_size 64
--learning_rate 2e-5
--num_train_epochs 3
--output_dir tmp/SST2_epoch3_lr2e-5_bs32_normc1.0_eps1_8_eps2_2.2_noise_layer_10_proj_dim_256
--overwrite_output_dir
--logging_steps 100
--save_steps 8000
--epsilon 8.0
--epsilon2 2.2
--noise_position encoder
--norm_c 1.0
--noise_layer 10
--encoder_sub_noise_position add_and_norm_2
--add_noise_inference True
--proj_dim -1
--local_dp False
--label_dp

Membership Inference Attacks

python attack/membership_inference_attack.py \
--model_name_or_path tmp/SST2_epoch3_lr2e-5_bs32_normc1.0_eps1_8_eps2_2.2_noise_layer_10_proj_dim_256 \
--task_name sst2 \
--max_seq_length 128 \
--per_device_eval_batch_size 64 \
--overwrite_cache

Sensitive Attribution Attack

python attack/sensitive_attributes_attack.py --model_name_or_path tmp/SST2_epoch3_lr2e-5_bs32_normc1.0_eps1_8_eps2_2.2_noise_layer_10_proj_dim_256

Embedding Inversion Attack

python attack/sensitive_attributes_attack.py --model_name_or_path tmp/SST2_epoch3_lr2e-5_bs32_normc1.0_eps1_8_eps2_2.2_noise_layer_10_proj_dim_256