Exploring Cross-Stage Adversarial Transferability in Class-Incremental Continual Learning (MMSP 2025)

Official repository of CSAT | Paper Link

Jungwoo Kim, Jong-Seok Lee

Summary

This paper investigates stage-transferred attacks in class-incremental continual learning (CSAT), showing that adversarial examples generated from earlier-stage models remain effective against later-stage models.

Environmental Set-up

conda create -n csat python=3.10
conda activate csat
git clone https://github.com/kjungwoo03/CSAT.git
cd CSAT
pip install -r requirements.txt

Quick Start

1. Train baseline model

The trained baseline models will be saved in the ./pretrained_models/{dataset}.

python train.py --dataset cifar100 --method gdumb --seed 42

2. Stage-transferred attacks

python attack.py --dataset cifar100 --method gdumb --seed 42

3. Others

For additional evaluations (e.g. defense, model similarity, complexity, etc.), see ./utils folder.

How to cite

TBD.