Killing It With Zero-Shot: Adversarially Robust Novelty Detection

---

Overview

This repository contains the official implementation of the paper "Killing It With Zero-Shot: Adversarially Robust Novelty Detection", presented at the ICASSP 2024 conference. The paper introduces a novel approach to novelty detection (ND) by leveraging adversarial robustness and robust features extracted from pre-trained models. Our method significantly improves the performance of ND under adversarial conditions, bridging the gap between nearest-neighbor techniques and robust feature-based methods.

---

Features

• Zero-Shot Novelty Detection: Leverages pre-trained models for robust anomaly detection without task-specific fine-tuning.
• Adversarial Robustness: Provides resistance against adversarial attacks such as PGD and FGSM.
• Multi-Testing Modes:
  • Anomaly Detection (AD)
  • Open Set Recognition (OSR)
  • Out-of-Distribution Detection (OOD)
• Configurable Testing: Customizable adversarial attacks, backbone models, datasets, and parameters.

---

Getting Started

1. Clone the repository:

   git clone https://github.com/Mohammadjafari80/ZARND.git
   cd ZARND

2. Install dependencies:

   pip install -r requirements.txt

3. Prepare datasets: Ensure datasets are downloaded and paths are updated accordingly (e.g., ~/cifar10 for CIFAR-10).

4. Run the code:

   python main.py --source_dataset cifar10 --label 0 --backbone resnet18_linf_eps8.0 --test_type ad --test_attacks PGD-10 --eps 4/255

---

Arguments

Argument	Default Value	Description
--source_dataset	cifar10	Source dataset for training and testing.
--source_dataset_path	~/cifar10	Path to the source dataset.
--target_dataset	None	Target dataset (only used for OOD testing).
--target_dataset_path	~/cifar100	Path to the target dataset (only used for OOD testing).
--model_path	./pretrained_models/	Path to pre-trained models.
--label	None	The class to consider as "normal" for anomaly detection. Must be specified for AD tests.
--eps	4/255	Perturbation limit for adversarial attacks.
--test_type	ad	Type of test to perform. Options: ad (Anomaly Detection), osr (Open Set Recognition), ood (Out-of-Distribution Detection).
--batch_size	128	Batch size for data loading.
--backbone	18	Backbone model to use. Options include ResNet variants with different robustness (e.g., resnet18_linf_eps2.0, wide_resnet50_2_linf_eps4.0) or standard ResNets (18, 50).
--test_attacks	None	List of adversarial attacks to test. Options: PGD-n (PGD with n steps), PGDA-n (advanced PGD with n steps), or FGSM.

---

Adversarial Testing

The adversarial tests are configurable using the --test_attacks argument. Supported attacks:

• FGSM: Fast Gradient Sign Method.
• PGD-n: Projected Gradient Descent with n steps.
• PGDA-n: Advanced PGD with n steps.

For example, to test with FGSM and PGD with 10 steps:

python main.py --source_dataset cifar10 --label 0 --test_attacks FGSM PGD-10

---

Results

Results are saved in the ./results/<test_type>/ directory. Filenames include dataset, backbone, and test type for easy identification.

---

Citation

If you find this code useful, please cite our work:

@INPROCEEDINGS{10446155,
  author={Mirzaei, Hossein and Jafari, Mohammad and Dehbashi, Hamid Reza and Sadat Taghavi, Zeinab and Sabokrou, Mohammad and Rohban, Mohammad Hossein},
  booktitle={ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)}, 
  title={Killing It With Zero-Shot: Adversarially Robust Novelty Detection}, 
  year={2024},
  pages={7415-7419},
  keywords={anomaly detection; adversarial robustness; zero-shot learning},
  doi={10.1109/ICASSP48485.2024.10446155}
}

---

Feel free to raise an issue or contribute to this repository!