The code accompanying our NeurIPS 2022 paper: LAMP: Extracting Text from Gradients with Language Model Priors.

For a brief overview, check out our blogpost.

• Install Anaconda.
• Create the conda environment:
  

conda env create -f environment.yml

• Enable the created environment:
  

conda activate lamp

• Download required files:
  

wget -r -np -R "index.html*" https://files.sri.inf.ethz.ch/lamp/
mv files.sri.inf.ethz.ch/lamp/* ./
rm -rf files.sri.inf.ethz.ch

• DATASET - the dataset to use. Must be one of cola, sst2, rotten_tomatoes.
• BERT_PATH - the language model to attack. Must be one of bert-base-uncased, huawei-noah/TinyBERT_General_6L_768D, models/bert-base-finetuned-cola, models/bert-base-finetuned-sst2, models/bert-base-finetuned-rottentomatoes for BERT_BASE, TinyBERT₆, and each of the three fine-tuned BERT_BASE-FT models (on each of the datasets).

• To run the experiment on LAMP with cosine loss:
  

./lamp_cos.sh BERT_PATH DATASET 1

• To run the experiment on LAMP with cosine loss on BERT_LARGE:
  

./lamp_cos_large.sh DATASET 1

• To run the experiment on LAMP with L1+L2 loss:
  

./lamp_l1l2.sh BERT_PATH DATASET 1

• To run the experiment on LAMP with L1+L2 loss on BERT_LARGE:
  

./lamp_l1l2_large.sh DATASET 1

• To run the experiment on TAG:
  

./tag.sh BERT_PATH DATASET 1

• To run the experiment on TAG on BERT_LARGE:
  

./tag_large.sh DATASET 1

• To run the experiment on DLG:
  

./dlg.sh BERT_PATH DATASET 1

• To run the experiment on DLG on BERT_LARGE:
  

./dlg_large.sh DATASET 1

• DATASET - the dataset to use. Must be one of cola, sst2, rotten_tomatoes.
• BATCH_SIZE - the batch size to use e.g 4.

• To run the experiment on LAMP with cosine loss:
  

./lamp_cos.sh bert-base-uncased DATASET BATCH_SIZE

• To run the experiment on LAMP with L1+L2 loss:
  

./lamp_l1l2.sh bert-base-uncased DATASET BATCH_SIZE

• To run the experiment on TAG:
  

./tag.sh bert-base-uncased DATASET BATCH_SIZE

• To run the experiment on DLG:
  

./dlg.sh bert-base-uncased DATASET BATCH_SIZE

• DATASET - the dataset to use. Must be one of cola, sst2, rotten_tomatoes.

• To run the ablation experiments in Table 4:
  

./ablation.sh DATASET

• SIGMA - the amount of Gaussian noise with which to defend e.g 0.001.

• To run the experiment on LAMP with cosine loss:
  

./lamp_cos.sh bert-base-uncased cola 1 --defense_noise SIGMA

• To run the experiment on LAMP with L1+L2 loss:
  

./lamp_l1l2.sh bert-base-uncased cola 1 --defense_noise SIGMA

• To run the experiment on TAG:
  

./tag.sh bert-base-uncased cola 1 --defense_noise SIGMA

• To run the experiment on DLG:
  

./dlg.sh bert-base-uncased cola 1 --defense_noise SIGMA

• ZEROED - the ratio of zeroed out gradient entries e.g 0.75.

• To run the experiment on LAMP with cosine loss:
  

./lamp_cos.sh bert-base-uncased cola 1 --defense_pct_mask ZEROED

• To run the experiment on LAMP with L1+L2 loss:
  

./lamp_l1l2.sh bert-base-uncased cola 1 --defense_pct_mask ZEROED

• To run the experiment on TAG:
  

./tag.sh bert-base-uncased cola 1 --defense_pct_mask ZEROED

• To run the experiment on DLG:
  

./dlg.sh bert-base-uncased cola 1 --defense_pct_mask ZEROED

• DATASET - the dataset to use. Must be one of cola, sst2, rotten_tomatoes.
• SIGMA - the amount of Gaussian noise with which to train e.g 0.001. To train without defense set to 0.0.
• NUM_EPOCHS - for how many epochs to train e.g 2.

• To train your own network:
  

python3 train.py --dataset DATASET --batch_size 32 --noise SIGMA --num_epochs NUM_EPOCHS --save_every 100

The models are stored under finetune/DATASET/noise_SIGMA/STEPS

@inproceedings{
    balunovic2022lamp,
    title={{LAMP}: Extracting Text from Gradients with Language Model Priors},
    author={Mislav Balunovic and Dimitar Iliev Dimitrov and Nikola Jovanovi{\'c} and Martin Vechev},
    booktitle={Advances in Neural Information Processing Systems},
    editor={Alice H. Oh and Alekh Agarwal and Danielle Belgrave and Kyunghyun Cho},
    year={2022},
    url={https://openreview.net/forum?id=6iqd9JAVR1z}
}