SPD: Shallow Backdoor Protecting Deep Backdoor Against Backdoor Detection

This repository provides the official implementation of our paper "SPD: Shallow Backdoor Protecting Deep Backdoor Against Backdoor Detection". This code is released for academic and research purposes only. Misuse of backdoor techniques may lead to serious ethical and legal consequences.

🔥 Overview

Backdoor attacks pose serious threats to the security of Deep Neural Networks (DNNs). However, most existing backdoor attacks are vulnerable to backdoor detection methods and human visual inspection. To address this, we propose a backdoor attack method named SPD (Shallow Protecting Deep). SPD implants:

• A deep backdoor in the frequency domain, which is imperceptible and encoded using an autoencoder.
• A shallow backdoor in the pixel domain, which distracts the defender and protects the deep backdoor.

🚀 Usage

1️⃣ Step 1: Prepare Data

Before launching the attack, make sure you have downloaded the datasets (CIFAR-10, GTSRB, ImageNet-12). You may modify the dataloader scripts if your dataset path is customized:

- dataloader_cifar.py
- dataloader_gtsbr.py
- dataloader_imagenet12.py

2️⃣ Step 2: Inject Backdoor

Run the following commands to execute SPD:

python spd_cifar10.py #CIFAR-10
python spd_gtsbr.py #GTSBR
python spd_imagenet12.py #ImageNet-12

🙏 Acknowledgements

We sincerely thank the authors and contributors of previous works in the field of backdoor attacks and defenses. In particular, we would like to acknowledge BackdoorBench (https://github.com/SCLBD/BackdoorBench).