ActiveChooN / adversarial-attack-sirius

Adversarial attack research project for Sirius and Yandex workshop

Adversarial attack Sirius and Yandex workshop project

An attack on image classifiers introduces unpredictable behavior for many computer vision systems. The goal of this project is to investigate a new type of attack, in which completely different images have the same representation in the neurogenic network.

References

• Excessive Invariance Causes Adversarial Vulnerability

Dataset

• MNIST&ImageNet

Train model

For training model run:

python3 main.py --model model_name --dataset dataset_name

• Available value for model is example.
• Available value for dataset is MNIST.
• Available value for optimizer is SGD.
• Use --train-batch-size, --test-batch-size, --epochs and --log-interval as training/testing settings.
• Use --lr and --momentum as optimizer settings.
• Use --save-model and --save-path for saving model in the desired path.
• Use the --no-cuda flag to train on the CPU rather than the GPU through CUDA.
• Use --log-level for setting logger level
• Use --metrics for different metrics set for the evaluator
• Set COMETML_API_KEY env variable for sending data to comet.ml

Authors

• Alex Babushkin
• Diana Zagidullina
• Dmitry Kalinin