Code for Poisoning Bayesian Inference via Data Deletion and Replication

This repository contains code for reproducing all the experiments in the paper "Poisoning Bayesian Inference via Data Deletion and Replication" (https://arxiv.org/abs/2503.04480).

Installation and dependencies

Install all required dependencies using

conda env create -f poisoning_bayesian_inference.yml

Finally, install package using

pip install -e .

Replication of the simulation study for linear regression

The results of the simulation study of Section 6.1 can be replicated with the notebook simulation_study_NIG.ipynb.

The additional results obtained with the same model presented in the supplementary materials can be replicated with:

• simulation_study_NIG_roudning_effect.ipynb for the continuous relaxation problem (Section 5.2 of supplementary materials).

• simulation_study_NIG_KL_evolution_heuristics.ipynb for the convergence speed of different heuristics (Section 5.3 of supplementary materials).

• simulation_study_NIG_noise_level_effect.ipynb for the noise level experiment (Section 5.4 of supplementary materials).

• simulation_study_NIG_decrease_uncertainty.ipynb and simulation_study_NIG_increase_uncertainty.ipynb for the uncertainty experiment (Section 5.5 of supplementary materials).

Replication of the linear regression experiments on real data

The attack on the house prices dataset from Section 6.2 can be replicated with the notebook house_prices.ipynb, and the one against the Mexico microcredit dataset with microcredit.ipynb (Section 6.3). The experiments with the gray-box attack scenario from Section 6.1.3 of supplementary materials can be found in the notebook house_prices.ipynb.

Replication of the logistic regression experiments for spam classification

The attack on the spam classifier from Section 6.3 of supplementary materials can be replicated with the notebooks spam_classification_send.ipynb and spam_classification_your.ipynb.