This is responsory is modified from the repo of the paper "Exploring Vulnerabilities in Spiking Neural Networks: Direct Adversarial Attacks on Raw Event Data", from ECCV 2024.

Requirements

torch ==  2.0.0+cu118
torchvision == 0.15.1+cu118
spikingjelly == 0.0.0.0.14
snntorch == 0.6.4
wandb == 0.16.1
numpy == 1.23.5

To install all the requirements, run:

pip install -r requirements.txt

File Structure

• data/: contains the datasets used in the paper.
• model/: contains the models used in the paper.
• attacks/: contains the attacks used in the paper.
• config/: the configuration file to set the parameters.
• utils/: the utils folder.
• main.py: the main file to run the attack.
• attacker.py: the attacker class files.
• requirements.txt: the requirements file.
• train.py: train code.
• README.md: this file.

Training

Use train.py and pass dataset.frame_number={x-y-z} to control the desired latencies.

Attack

Use main.py and pass the same argument as above. If the model is trained on multiple latencies, then the Self-Ensemble defense is conducted automatically after each successful attack (on single latency). Please use attack.init_alpha_mode and attack.add_position_label_mode to control the attacker's behavior.