Rethinking the Privacy of Text Embeddings: A Reproducibility Study of “Text Embeddings Reveal (Almost) As Much As Text”

Vec2Text Reproducibility Study

This repository contains code for reproducing results from the paper "Text Embeddings Reveal (Almost) As Much As Text".

Abstract: This work investigates the Vec2Text method, which frames the embedding inversion problem as a controlled generation task. Our study aims to validate Vec2Text's ability to reconstruct text from embeddings, highlighting the privacy risks associated with embedding inversion. We reproduce Vec2Text's performance in both in-domain and out-of-domain settings, verifying its effectiveness while noting some discrepancies due to experimental details. We extend the study by exploring parameter sensitivity, password reconstruction, and embedding quantization as a defense strategy. The study concludes with insights into Vec2Text's robustness and potential risks, emphasizing the need for further research on embedding inversion methods and strategies to prevent them.

Project structure

We use the following project structure:

src/ - source directory
notebooks/ - jupyter notebooks
runs/ - configuration files for running experiments.
jobs/ - sbatch jobs for running experiments
scripts/ - scripts for running experiments
tests/ - tests for the project
docs/ - documentation for some of the findings

Installation

1. Clone the repository
2. Setup correct Python environment using Conda environment:

conda env create -f environment.yaml
conda activate vec2text-repro

Alternativelly install Python using pyenv and then install dependencies using pip:

pyenv install 3.11.6
pyenv local 3.11.6
pip install -r requirements.txt

Usage

To run inference scripts, you need to first login to wandb:

wandb login

Then you can run the scripts:

python scripts/inference.py <RUN_CONFIG>

where <RUN_CONFIG> is the path to the run config file. All of the config files are located in the runs directory.

For example, to run quantization experiments, you can run:

python scripts/inversion_attacks.py runs/attack_quant.yaml

Runs and experiment setups

The runs are organized in the runs directory. Each run is a YAML file that contains the configuration for the experiment. The runs are organized by the type of experiment:

• table_1 - Table 1 (In-domain reproduction performance of Vec2Text)
• table_2 - Table 2 (Out-of-domain reproduction performance)
• attack - Figure 3 (Noise injection experiments)
• passwords_check - Table 4 (Password reconstruction experiments)
• attack_quant - Table 5 (Weight quantization experiments)
• seq_len_influence - Figure 2 (Sequence length influence experiments)
• table_1_extension - Figures 4 and 5 (Extended experiments on vec2text parameters)

Configuration

• model_name - name (path) of the encoder model to use
• corrector_name - name (path) of the corrector model to use
• dataset - name of the dataset to use (e.g. quora)
• batch_size - batch size for inference
• num_steps - number of steps to run while correcting the embedding inversion
• add_gaussian_noise - (default false) whether to add Gaussian noise to the embeddings
• noise_mean - (not used if add_gaussian_noise is set to false) mean of the Gaussian noise
• noise_std - (not used if add_gaussian_noise is set to false) standard deviation of the Gaussian noise
• noise_lambda(not used if add_gaussian_noise is set to false) scalar for Gaussian noise