Disparate Privacy Vulnerability: Targeted Attribute Inference Attacks and Defenses

The accompanying artifacts provide a comprehensive framework for understanding and replicating the experiments described in our paper, Disparate Privacy Vulnerability: Targeted Attribute Inference Attacks and Defenses. These artifacts are designed to support open science practices, ensuring reproducibility and transparency in research. They include datasets, source code, and Jupyter notebooks used to implement and evaluate the proposed attacks and defenses. The datasets encompass both publicly available and partitioned subsets to replicate training and testing scenarios. The codebase is modular, with dedicated scripts for data preprocessing, model setup, experiment configuration, and attack implementation. Additionally, the notebooks encapsulate the step-by-step execution of the experiments, enabling users to validate key results such as the relationship between correlation and attack performance, the efficacy of disparity inference and targeted inference attacks, and the impact of proposed defenses like BCorr. These resources are shared in compliance with the conference’s open science policy and hosted on a platform ensuring permanent access. Environment Setup: To set up the environment, create a virtual environment using Conda or Python’s venv, then activate it. Install dependencies with pip install -r requirements.txt. To run Jupyter notebooks, install notebook using pip install notebook. Once done, the environment is ready for running the provided scripts and notebooks. Please extract the zip files <PATH_TO_DATASET>.zip and <PATH_TO_MODEL>.zip Datasets: census19.csv contains the Census-19 dataset. Adult_35222.csv and Adult_10000.csv contains the Adult dataset partitions for training and testing respectively. texas_100_cleaned.csv contains the preprocessed version of the Texas-100X dataset used in the experiments Codebase: Source Files: data_utils.py contains code to load dataset and preprocess data. It also contains code to sample data matching target correlation at a subgroup level as described in section 6.1 paragraph 'Sampling Technique'. model_utils.py contains code to define the model architecture and training hyperparameters. experiment_utils.py contains code to setup experiment for a particular scenario. attack_utils.py contains the implementation of existing attacks including CSMIA, LOMIA, imputation attack, and Neuron Importance Attack. whitebox_attack.py contains helper functions needed to perform the neuron importance attack. disparity_inference_utils.py contains the implementation of Confidence Matrix generation (Algorithm 1), Angular Difference computation (Algorithm 2). targeted_inference.py contains the implementation of the targeted attribute inference attacks (section 5.3). bcorr_utils.py contains the implementation of the sampling stage of the BCorr defense (section 7.2). Notebooks: correlation_vs_attack_performance.ipynb contains the code to run the experiment described in section 4.1 which shows the strong connection between correlation and attack performance. angular_difference_by_sex.ipynb contains the code to run the experiment described in section 4.2 which shows how the angular difference can be used to identify vulnerable groups. imputation_vs_ai_aux_size_and_distrib_diff.ipynb contains the code to run the experiment described in section 6.2, which compares the performance between ideal imputation and practical imputation attacks. disparity_inference.ipynb contains the code to run the experiment described in section 6.3 which shows how the disparity inference attack can be used to rank groups based on their vulnerability. targeted_attribute_inference.ipynb contains the code to run the experiments in section 6.4 which shows how the targeted attribute inference attack outperforms their untargeted counterparts and practical imputation attacks. mutual_info_reg.ipynb contains the code to run the experiments in section 7.1 which shows the ineffectiveness of existing defenses in mitigating disparate vulnerability. bcorr_defense.ipynb contains the code to run the experiments in section 7.2 which shows how the BCorr defense can be used to mitigate the vulnerability of the model to the targeted attribute inference attack.