Membership Inference Attacks on Seq-to-Seq Models

Installation and Requirements

• System Requirements : TensorFlow 2.1+, Python 3.5+ Installation

1. Create and activate a Virtual environment

python3 -m virtualenv venv source venv/bin/activate

2. Install dependencies

pip install -r requirements.txt

Running the attack

To train machine translation target model

python main.py

To test attack 1 (Using average rank thresholding)

python attack1.py

To test attack 2 (Shadow models - rank histogram)

python attack2.py

Using the tool on an existing model

In progress

File Structure

In progress

TODOs

• Restructure tool
• Check if attacks are correctly implemented
• Make usage of other format datasets easier
• Comments, documentation
• etc.

Misc

Files adapted from https://github.com/csong27/auditing-text-generation:

• helper.py: defines custom architecture for user level NMT
• load_sated.py1: dataloading from SATED dataset
• sated_nmt.py: trains target and shadow models for SATED dataset
• sated_nmt_ranks.py: gets ranks for target and shadow models