/ted

Primary LanguageJupyter Notebook

Robust Backdoor Detection for Deep Learning via Topological Evolution Dynamics

This is the official repository for the paper "Robust Backdoor Detection for Deep Learning via Topological Evolution Dynamics" presented at IEEE Symposium on Security and Privacy (S&P) 2024.

Topology Persistence Digram

Source-Specific and Dynamic-Triggers (SSDT) Attack

To execute the Source-Specific and Dynamic-Triggers (SSDT) attack on the CIFAR-10, MNIST, or GTSRB dataset, use the following configuration:

  • Command: python train_SSDT.py
  • Arguments:
    • --dataset [cifar10/mnist/gtsrb] (replace with the desired dataset)
    • --attack_mode SSDT
    • --n_iters 300

Example command for CIFAR-10:

python train_SSDT.py --dataset cifar10 --attack_mode SSDT --n_iters 300

Topological Evolution Dynamics (TED) Defense

To explore the TED defense methodology, use the TED.ipynb Jupyter Notebook provided in this repository.