This repository contains the source code to support the paper The Hammer and the Nut: Is Bilevel Optimization Really Needed to Poison Linear Classifiers?, by Antonio Emanuele CinĂ , Sebastiano Vascon, Ambra Demontis, Battista Biggio, Fabio Roli and Marcello Pelillo, accepted at IJCNN 2021.
Please use the command
$ conda env create -f environment.yml
to create the conda environment beta_poison
.
Then you need to activate the conda env by running conda activate beta_poison
The file run_experiments.sh
contains a list of commands to replicate the experiments and results
proposed in our paper. Output files are, by default, saved in "IJCNN_Experiments/" dir.
Experiments for MNIST Pairs (4 vs. 0 and 9 vs. 8):
$ ./run_experiments.sh mnist_bin
Experiments for CIFAR-10 Pairs (frog vs. ship and frog vs. horse):
$ ./run_experiments.sh cifar_bin
Experiments for MNIST triplet are obtained with:
$ ./run_experiments.sh mnist_triplet
Ablation study evaluates the effect of k (number of prototypes) during the optimization procedure. To run it use:
$ ./run_experiments.sh mnist_ablation # ablation on mnist
$ ./run_experiments.sh cifar_ablation # ablation on cifar
- antonioemanuele.cina [at] unive [dot] it