This is the pytorch implementation for our ICDE 2022 paper:
Rong, Dazhong, et al. "FedRecAttack: Model Poisoning Attack to Federated Recommendation." 2022 IEEE 38th International Conference on Data Engineering (ICDE). IEEE, 2022. Paper in arXiv
- Python 3.7.9
- numpy==1.18.5
- torch==1.7.0+cu101
To run FedRecAttack on MovieLens-100K with :
python main.py --dataset=ml-100k/ --attack=FedRecAttack --clients_limit=0.05 --items_limit=60 --part_percent=1
There are three choices on dataset:
--dataset=ml-100k/, --dataset=ml-1m/ and --dataset=steam/.
There are four choices on attack:
--attack=FedRecAttack, --attack=Random, --attack=Bandwagon and --attack=Popular.
Arguments: attack=FedRecAttack,dim=32,path=Data/,dataset=ml-100k/,device=cuda,lr=0.01,epochs=200,batch_size=256,grad_limit=1.0,clients_limit=0.05,items_limit=60,part_percent=1,attack_lr=0.01,attack_batch_size=256
Load data done [5.4 s]. #user=990, #item=1682, #train=99056, #test=943
Target items: [894].
output format: ({Sampled HR@10}), ({ER@5},{ER@10},{NDCG@10})
Iteration 0(init), (0.0848) on test, (0.0021, 0.0054, 0.0025) on target. [1.1s]
Iteration 1, loss = 72.84028 [3.2s], (0.0732) on test, (0.3869, 0.4019, 0.3799) on target. [1.0s]
Iteration 2, loss = 72.82960 [2.0s], (0.0764) on test, (0.5198, 0.5220, 0.5133) on target. [1.1s]
Iteration 3, loss = 72.82328 [2.0s], (0.0817) on test, (0.6045, 0.6077, 0.6001) on target. [1.1s]
Iteration 4, loss = 72.81755 [2.0s], (0.1166) on test, (0.7170, 0.7192, 0.7155) on target. [1.0s]
Iteration 5, loss = 72.80792 [2.0s], (0.2174) on test, (0.7814, 0.7824, 0.7809) on target. [1.0s]
... ...
The codes are for learning and research purposes only.