Revisiting Adversarially Learned Injection Attacks Against Recommender Systems

A PyTorch implementation of paper:

Revisiting Adversarially Learned Injection Attacks Against Recommender Systems, Jiaxi Tang, Hongyi Wen and Ke Wang , RecSys '20

Requirements

• Python 3.5+
• Please check requirements.txt for required python packages.

Running experiment

Synthetic dataset

synthetic_exp.ipynb includes a self-contained implementation for:

1. Generating synthetic data
2. Experiments in the paper

Real-world dataset

Please refer to the following steps to run experiments on real-world dataset (i.e., Gowalla):

1. Install required packages.
2. Create a folder for experiment outputs (e.g., logs, model checkpoints, etc)

   cd revisit_adv_rec
   mkdir outputs
   

3. To generate fake data for attacking, change the configs in generate_attack_args.py (or leave as it is) then run:

   python generate_attack.py
   

   You will find the fake data stored in the outputs/ folder, such as outputs/Sur-ItemAE_fake_data_best.npz
4. To inject fake data and evaluate the recommender performance after attack, modify the configs in evaluate_attack_args.py (you need to point the fake_data_path to your own) then run:

   python evaluate_attack.py
   

5. To evaluate each victim model's performance without fake data (i.e., Without attack in Figure 5(a)), set fake_data_path=None in evaluate_attack_args.py then run:

   python evaluate_attack.py
   

Below are the logs obtained from using WRMF+SGD method for attack:

• gen_attack_wrmf_sgd
• eval_attack_clean
• eval_attack_wrmf_sgd

Citation

If you use the code in your paper, please cite the paper:

@inproceedings{tang2020revisit,
  title={Revisiting Adversarially Learned Injection Attacks Against Recommender Systems},
  author={Tang, Jiaxi and Wen, Hongyi and Wang, Ke},
  booktitle={ACM Conference on Recommender Systems},
  year={2020}
}