Targeted Adversarial Attack against Multimedia Recommender Systems (TAaMR)

GitHub repository of the DMSL2020 paper: Targeted Adversarial Attack against Multimedia Recommender Systems, published by Tommaso Di Noia, Daniele Malitesta and Felice Antonio Merra.

Paper available at Sisinflab publications web page.

The architectural overview of the proposed approach is as below:

Data preparation

To run the experimental section it is necessary to:

Create in ./data/ a directory to store the dataset.
Move in ./data/<dataset_name>/.
Insert train and test file user_id\titem_id (users' and items indices need to be from 0 to N-1).
Create directory ./data/<dataset_name>/original_images/images/.
Store in ./data/<dataset_name>/original_images/images/ all the images using the same item_id in the format item_id.jpg.

Experiments

Operations to be executed (in src/):

classify_extract.py - Executes the classification and the feature extraction on original images (the files classes.csv and features.npy are stored in ./data/<dataset_name>/original_images/).
rec_generator.py - Generates the recommendation for the clean classes.
results_analyzer.py - Evaluates the recommendation lists useful for select the origin and target class for the targeted adversarial attack.

Results evaluation

For the generation of each attack, run the script classify_extract_attack.py. It will create a new directory named ./data/<dataset_name>/<attack_name_parameters>/.

After each attack, run again rec_generator.py and results_analyzer.py to generate the new recommendations and evaluate these results. Additionally, run evaluate_visual_images.py to evaluate the visual metrics on the attacked images.

Implemented attacks

FGSM
PGD

Requirements

Python 3.6.8
CUDA 10.1
cuDNN 7.6.4

All the requirements are in the file requirements.txt

pip install -r requirements.txt

If you use this code, please cite us: