recad 0.0.2

A unified framework for recommender system attacking

RecAD

A unified library for recommender attack and defense

RecAD is a unified library aiming at establishing an open benchmark for recommender attack and defense. With a few line of codes, you can quickly construct a attacking pipeline.

The supported modules currently include:

• Datasets: ml1m, yelp, Amazon-game, Epinions, Book-crossing, BeerAdvocate, dianping, food, ModCloth, ratebeer, RentTheRunway. Please checkout more details in data/
• Victim Models: MF, LightGCN, NCF.
• Attack Models: Heuristic(random, average, segment, bandwagon); AUSH; AIA; Legup

🚀🚀🚀 We are opening for any contribution or suggestion about adding more datasets and models

Install

Install by pip:

pip install recad

Or from source:

git clone https://github.com/gusye1234/recad.git
cd recad
pip install -e "."

Quick Start

Try it from command line:

recad_runner --attack="aush" --victim="lightgcn"

Or you can write your own script:

from recad import dataset, model, workflow

dataset_name = "ml1m"
config = {
    # quickly asscess the dataset with implicit feedback
    "victim_data": dataset.from_config("implicit", dataset_name, need_graph=True),
    # sample part of the explicit dataset as the attack data
    "attack_data": dataset.from_config("explicit", dataset_name).partial_sample(
        user_ratio=0.2
    ),
    # set up models config, and later will be instantiated in workflow
    "victim": model.from_config("victim", "lightgcn"),
    "attacker": model.from_config("attacker", "aush"),
    "rec_epoch": 20,
}
workflow_inst = workflow.from_config("no defense", **config)
# run the attacking
workflow_inst.execute()

Docs

recad is designed to help users use and debug interactively, and mainly has three modules: dataset, model, workflow

Dataset

import recad

# how many datasets we support?
print(recad.print_datasets())

# how many configs for one dataset?
recad.dataset_print_help("ml1m")

# init from a dataset with default parameters
dataset = recad.dataset.from_config("implicit", "ml1m")

# init from a dataset and modifies parameters
dataset = recad.dataset.from_config("implicit", "ml1m", test_batch_size=50, device="cuda")

Model

# how many models we support?
print(recad.print_models())

#how many configs for one model?
recad.model_print_help("lightgcn")
  
# lazy-init from a model with default parameters
# Not a torch.nn.Module class! Can't be used for training and inferring
victim_model = recad.model.from_config("victim", "lightgcn")

# lazy-init from a model and modifies parameters
# Not a torch.nn.Module class! Can't be used for training and inferring
victim_model = recad.model.from_config("victim", "lightgcn", latent_dim_rec=256, lightGCN_n_layers=2)

# Model's have some parameters that are related to some runtime module, e.g. dataset
# `victim_model` now is an actually torch.nn.Module 
dataset = recad.dataset.from_config("implicit", "ml1m")
victim_model = recad.model.from_config("victim", "lightgcn", dataset=dataset).I()

Workflow

# how many workflows we support?
print(recad.print_workflows())

#how many configs for one workflow?
recad.workflow_print_help("no defense")

# init a workflow takes all the components we mentioned before
config = {
        "victim_data": ..., # Your dataset for victim model, using dataset.from_config...
        "attack_data": ..., # Your dataset for attacker model, using dataset.from_config...
        "victim": ..., # Your victim model, using model.from_config...
        "attacker": model.from_config(
            "attacker", ARG.attack, filler_num=ARG.filler_num
        ), # Your attacker model, using model.from_config...
        "rec_epoch": ..., # Your training epoch for victim model, Int
        "attack_epoch": ..., # Your training epoch for attacker model, Int
    }
workflow_inst = workflow.from_config("no defense", **config)

# Start
workflow_inst.execute

Please checkout the whole pipeline and more details for each module in recad.main🤗.

Confused about what a component is doing? Each component instance in recad will have a print_help method to return the input/output information:

dataset.print_help()
# (ml1m)Information:
# ╒════════════════════╤═══════════════════════════════════╕
# │ n_users            │ 5950                              │
# ╞════════════════════╪═══════════════════════════════════╡
# │ n_items            │ 3702                              │
# ├────────────────────┼───────────────────────────────────┤
# │ train_interactions │ 468649                            │
# ├────────────────────┼───────────────────────────────────┤
# │ valid_interactions │ 49390                             │
# ├────────────────────┼───────────────────────────────────┤
# │ test_interactions  │ 49494                             │
# ├────────────────────┼───────────────────────────────────┤
# │ train_dict         │ <class 'dict'> 5950               │
# ├────────────────────┼───────────────────────────────────┤
# │ valid_dict         │ <class 'dict'> 5583               │
# ├────────────────────┼───────────────────────────────────┤
# │ test_dict          │ <class 'dict'> 5677               │
# ├────────────────────┼───────────────────────────────────┤
# │ graph              │ torch torch.float32, (9652, 9652) │
# ╘════════════════════╧═══════════════════════════════════╛
# (ml1m)Batch data:
# ╒════════════════╤═════════════╤═══════════════╕
# │ name           │ type        │ shape         │
# ╞════════════════╪═════════════╪═══════════════╡
# │ users          │ torch.int64 │ batch[0~2048] │
# ├────────────────┼─────────────┼───────────────┤
# │ positive_items │ torch.int64 │ batch[0~2048] │
# ├────────────────┼─────────────┼───────────────┤
# │ negative_items │ torch.int64 │ batch[0~2048] │
# ╘════════════════╧═════════════╧═══════════════╛

Contribution

Install pre-commit first to make sure the commits you made is well-formatted:

pip install pre-commit
pre-commit install