CryptoBackdoor

This repository contains the official code for the paper "Rethinking the Stealthiness of Cryptographically Undetectable Backdoors in Practical RFF Learning", accepted by KDD 2026.

Environment

conda create --name cryptobackdoor python=3.9
conda activate cryptobackdoor
pip install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/cu124
pip install numpy scikit-learn scipy matplotlib tqdm cvxpy

CIFAR-10 is downloaded automatically. For Covertype, please download and place the data file at:

./dataset/covtype.data

Directory

CryptoBackdoor/
├── README.md
├── detect.py
└── src/
    ├── cifar10_acc.py
    ├── cifar10_clwe.py
    ├── covertype_acc.py
    └── covertype_clwe.py

Files

src/*_acc.py

Compares accuracy under different preprocessing choices such as raw input, standardization, and scaling on CIFAR-10 and Covertype.

src/*_clwe.py

Runs the CLWE/Goldwasser-style backdoor attack on CIFAR-10 and Covertype. It trains an RFF classifier, constructs a malicious feature matrix and trigger vector bk, then reports clean accuracy, triggered accuracy, and prediction flip ratio.

detect.py

Evaluates a detector for triggered samples. It marks a sample as suspicious when its infinity norm exceeds a threshold and reports detection rate and false positive rate under different preprocessing settings.

Workflow

mkdir -p dataset figs

python src/cifar10_acc.py
python src/covertype_acc.py

python src/cifar10_clwe.py
python src/covertype_clwe.py

python detect.py