Calibrated Losses for Adversatial Robust Reject Option Classification

Description

This repository contains the implementation of algorithms to learn linear reject option classifiers using the Double Sigmoid loss and Double Ramp loss when the data is $\ell_{2}$-norm adversarially perturbed.

Installation

1. Clone the repository: https://github.com/Vrund0212/Calibrated-Losses-for-Adversarial-Robust-Reject-Option-Classification.git
2. Navigate into the directory: cd Calibrated-Losses-for-Adversarial-Robust-Reject-Option-Classification
3. Install dependencies: pip install -r environment.yml

Explanation

1. There are two models that can be trained, one using the Double Sigmoid loss (--dsl) and another using the Double Ramp loss (--drl).
2. $\beta$ (--beta) denotes the shift parameter and $\gamma$ (--gamma) is the radius of the perturbation ball.
3. Train $\gamma$ (--train_gamma) is the perturbation used during training phase whereas Test $\gamma$ (--test_gammas) are the perturbations at test time.
4. $d$ (--cost) is the cost of rejection, which is assumed to be constant for all examples.
5. $\mu$ (--mu) is the slope.

Usage

# Example command to run for model trained using Double Sigmoid loss
python -Wi main.py --model dsl --beta 0.5 --train_gamma 0.01 --test_gammas 0 0.01 0.1 --cost 0.2 --mu 1.2 --tqdm