Attacking Pre-trained Recommendation (APRec)

Yiqing Wu, Ruobing Xie, Zhang Zhao‡, Fuzhen Zhuang, Jie Zhou, Yongjun Xu, Qing He Attacking Pre-trained Recommendation In SIGIR2023 This is our implementation of the paper:

Please cite our SIGIR2023 paper if you use our codes. Thanks!*

@inproceedings{WuAtt2022,
author = {Wu, Yiqing and Xie, Ruobing and Zhang, Zhao and Zhu, Yongchun and Zhuang, Fuzhen and Zhou, Jie and Xu, Yongjun and He, Qing},
title = {Attacking Pre-Trained Recommendation},
year = {2023},
isbn = {9781450394086},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3539618.3591949},
doi = {10.1145/3539618.3591949},
keywords = {security, pre-trained recommendation, attack},
location = {Taipei, Taiwan},
series = {SIGIR '23}
}

Our implementation is based on the ReChorus

Example to run the codes

For global attacking:

On pre-training stage:

python main.py --model_name Attack4AliEC --emb_size 64 --num_layers 1 --num_heads 1 --lr 1e-4 --l2 1e-6 --history_max 100 --dataset 'AttackAliEC' --gpu 1 --stage 1 --RATE 0.0

--RATE means the replacement rate, -1.0 means pre-training a clean model

On downstream dataset:

without fine-tuning:

python main.py --model_name Attack4AliEC --emb_size 64 --num_layers 1 --num_heads 1 --lr 1e-4 --l2 1e-6 --history_max 100 --dataset 'AttackAliEC' --gpu 1 --stage 4 --RATE 0.0

fine-tuning:

python main.py --model_name Attack4AliEC --emb_size 64 --num_layers 1 --num_heads 1 --lr 1e-4 --l2 1e-6 --history_max 100 --dataset 'AttackAliEC' --gpu 1 --stage 3 --RATE 0.0

For target group Attacking & prompt tuning:

For BRA:

python main.py --model_name GroupAttack4AliEC --emb_size 64 --num_layers 1 --num_heads 1 --lr 1e-4 --l2 1e-6 --history_max 100 --dataset 'AttackAliEC' --gpu 1 --stage 1 --RATE 0.0

For 𝑃𝐸A : step 1 Training a clean model

python main.py --model_name GroupAttack4AliECPrompt --emb_size 64 --num_layers 1 --num_heads 1 --lr 1e-4 --l2 1e-6 --history_max 100 --dataset AttackAliEC --gpu 1 --stage 1  --RATE -1.0 --w_prompt 0

step 2 Training a prompt:

python main.py --model_name GroupAttack4AliECPrompt --emb_size 64 --num_layers 1 --num_heads 1 --lr 1e-4 --l2 1e-6 --history_max 100 --dataset AttackAliEC --gpu 1 --stage 2  --RATE -1.0 --w_prompt 1

step 3 Training an attacked model:

python main.py --model_name GroupAttack4AliECPrompt --emb_size 64 --num_layers 1 --num_heads 1 --lr 1e-4 --l2 1e-6 --history_max 100 --dataset AttackAliEC --gpu 1 --stage 5  --RATE 0.5 --w_prompt 1

step 4 Fine-tuning prompt on downstream:

python main.py --model_name GroupAttack4AliECPrompt --emb_size 64 --num_layers 1 --num_heads 1 --lr 1e-4 --l2 1e-6 --history_max 100 --dataset AttackAliEC --gpu 1 --stage 7  --RATE -1.0 --w_prompt 1