Hidden_adversarial attacks

We propose concealed adversarial attacks for different time-series models that are hard to detect visually or via a specially trained discriminator. To achieve this goal, we train a discriminator model to detect adversarial perturbation in data. This model allows to define an additional regularization for an attack. The proposed adversarial attack with the regularization makes attacks harder to detect. Our claims are supported by extensive benchmarking via datasets from UCR and different models: recurrent, convolution, state-space, and transformers. For our proposed metric, a mix of concealability and effectiveness, we find an improvement over vanilla attacks. In some cases, the findings are more intricate, as adversarial vulnerability significantly changes for different models and datasets. Thus, our proposed attack can bypass the discriminator, raising the challenge of creating neural network models for time-series data that can resist adversarial attacks.

Data

You need to download UCR dataset dataset and put it into the data/UCR directory.

Quick Start

If you want to work with project with Docker, you can use folder docker_scripts. Firstly copy file credentials_example to credentials and tune it with your variables for running docker. After you need to make docker image using command:

cd docker_scripts
bash build

For creating container run:

bash launch_container

All the requirements are listed in requirements.txt For install all packages run

pip install -r requirements.txt

After you need to create folders checkpoints for saving classifier weights and results for saving adversarial attacks results.

Where are three basic steps: train classifier, attack model, train discriminator. To run these steps you need to rename folder config_examples to config, then change assosiated config files in "config" folder and after that run assosiated python scrits train_classifier.py, attack_run.py and train_discriminator.py.

For example:

python train_classifier.py

Describtion

The goal of the project is to create hardly detected adversarial attacks for time-series models.

Content

File or Folder	Content
checkpoints	folders for saving weights of the models
config	folder contains config files with params of models and paths
data	folder contains datasets For LSTM model (Ford_A) and datasets UCR, UEA datasets for TS2Vec models
docker_scripts	folder for set environment in docker container (if needed)
notebooks	folder with notebooks for data visualisation and small experiments
src	folder with code

Configs structure

We use a hydra-based experiment configuration structure. To run train_discriminator.py and attack_run.py you need to set the following configs:

• - dataset: choose the dataset you need from the config/dataset/ directory.
• - model: choose the model for discriminator from the config/model/ directory.
• - model@attack_model: choose the classifier model from the config/model/ directory which was trained using a script train_classifier.py.
• - model@disc_model_check: select the model for discriminator from the config/model/ directory if you want to get metrics of another trained discriminator. Note that if you want to use it, set use_disc_check: True
• - attack_disc@attack: choose the type of adversarial attack from the config/attack/ directory.