Configs

The Configs directory contains various configuration files and settings used for attacks, defenses, and training in the backdoor learning system. This directory is organized into subdirectories and files, each serving a specific purpose.

Directory Structure

configs/
├── attacks/
│   ├── audio/
│   ├── image/
│   ├── text/
│   └── video/
├── defenses/
├── __init__.py
├── default.yaml
├── settings.py
├── train_audio_args.yaml
├── train_image_args.yaml
├── train_text_args.yaml
└── train_video_args.yaml

attacks

The attacks subdirectory contains configuration files specific to different types of data attacks, organized by data type: audio, image, text, and video. Each folder contains YAML configuration files defining parameters and settings for various backdoor attack methods in the respective data domains.

defenses

The defenses subdirectory includes configuration files for defense mechanisms against backdoor attacks. These files outline settings for different defense strategies to mitigate the effects of data poisoning in various data types.

Configuration Files

settings.py

The settings.py file defines key directories and lists of available attacks, targets, datasets, models, and defenses, as well as configurations like poison rates. This file serves as a central location for defining paths and lists for commonly used components in the system.

Directory Paths:

BASE_DIR: The base directory of the project.
DATA_DIR: Directory where data is stored.
LOG_DIR: Directory for logging information.
TEST_DIR: Directory for testing files.
POISON_DATA_DIR: Directory for poisoned data storage.

TYPES: Specifies supported data types for backdoor attacks, including image, text, and audio.

ATTACKS: Defines supported attack types for each data type:

image: Includes attack methods like badnet, blend, bpp, wanet, and others.
text: Includes attacks like badnet, addsent, and stylebkd.
audio: Includes attacks such as badnet, blend, gis, and ultrasonic.
video: Limited to badnet and tuap for video data.

TARGETS: Specifies target labels for each dataset used in attack setups, e.g., “jazz” for gtzan or 0 for voxceleb1identification.

DATASETS: Defines available datasets for each data type:

image: e.g., cifar10, gtsrb, celeba.
text: e.g., imdb, dbpedia, sst2.
audio: e.g., speechcommands, gtzan, voxceleb1identification.
video: e.g., hmdb51.

MODELS: Lists supported models for each data type, such as:

image: resnet18, vit_b_16.
text: bert, gpt2, roberta.
audio: audiocnn, lstm.
video: r3d.

DEFENSES: Contains a list of available defense mechanisms, including ac, strip, finetune, nc, onion, and others.

POISON_RATE_LST: Lists various poisoning rates to test, ranging from 0.000 to 0.5.

default.yaml

The default.yaml file defines the default settings for training and testing, including model configurations, dataset information, and other training parameters.

Key Parameters:

attack_label_trans: Type of label transformation for the attack, e.g., all2one.
attack_target: Target label for the attack, default is 0.
client_optimizer: Optimizer type, e.g., sgd.
dataset: Dataset to be used, e.g., cifar10.
frequency_save: Frequency to save model checkpoints.
batch_size: Batch size for training, default 128.
lr: Learning rate, default 0.01.
model: Model architecture, e.g., resnet18.
pratio: Poisoning ratio, default 0.1.
epochs: Number of training epochs, default 100.

train_audio_args.yaml, train_image_args.yaml, train_text_args.yaml, train_video_args.yaml

These YAML files provide specific configurations for training with different data types: audio, image, text, and video. Each file includes parameters such as model type, number of epochs, batch size, learning rate, and optimizer settings. They allow customized settings based on the data type being used in training.

model: Model architecture for each data type, e.g., audiocnn for audio.
epochs: Number of epochs for training.
batch_size: Batch size for training.
client_optimizer: Optimizer type, typically sgd.
lr: Learning rate.
weight_decay: Weight decay for regularization.