Data-Poison-Incremental-Learning

This is a PyTorch implementation of the continual learning experiments on adversirial envioriments.

Environment settings and libraries we used in our experiments

This project is tested under the following environment settings:

• Google Colab
• PyTorch: >= 1.6.0
• Torchvision: >= 0.6.0

Running commands

running on clean dataset

!python -W ignore adv_si.py --dataset=clean --lr=0.0001 --batch_size=128 --epochs=10 --damping=0.01 --importance=5
!python -W ignore adv_ewc.py --dataset=clean --lr=0.0001 --batch_size=128 --epochs=10 --online=True --ewc_lambda=5000
!python -W ignore adv_ewc.py --dataset=clean --lr=0.0001 --batch_size=128 --epochs=10 --ewc_lambda=5000

running on label flipping attacks

!python -W ignore adv_si.py --dataset=lf --lr=0.0001 --batch_size=128 --epochs=10 --percentage=10 --damping=0.01 --importance=5
!python -W ignore adv_ewc.py --dataset=lf --lr=0.0001 --batch_size=128 --epochs=10 --percentage=10 --online=True --ewc_lambda=5000
!python -W ignore adv_ewc.py --dataset=lf --lr=0.0001 --batch_size=128 --epochs=10 --percentage=10 --ewc_lambda=5000

running on our attacks

!python -W ignore adv_si.py --dataset=adv --lr 0.0001 --batch_size 128 --epochs 10 --percentage 5 --damping 0.01 --importance 5 --num_steps 240 --decay 1.0 --epsilon 0.1 --rule 'adaptive'
!python -W ignore adv_ewc.py --dataset=adv --lr 0.0001 --batch_size 128 --epochs 10 --percentage 5 --online --ewc_lambda 5000 --num_steps 250 --decay 1.0 --epsilon 0.1 --rule 'adaptive'
!python -W ignore adv_ewc.py --dataset=adv --lr 0.0001 --batch_size 128 --epochs 10 --percentage 20 --ewc_lambda 5000 --num_steps 250 --decay 1.0 --epsilon 0.1  --rule 'fixed'

Data

Please extract dataset to .npy file. It will save 10x time training time due to the limitation of IO. Also, the poison algorithm is build based on the .npy data. Extracted filese can be found here.