tensorflow

TensorFlow Detection Module

This is the TensorFlow implementation of the adversarial detection methods, migrated from the PyTorch version.

Installation

pip install -r requirements.txt

Usage

Training Models

To train a CNN model for a specific dataset:

python train_model.py -d mnist -e 50 -b 128
python train_model.py -d cifar -e 100 -b 512
python train_model.py -d svhn -e 100 -b 256

Generating Adversarial Examples

To generate adversarial examples:

python generate_adv.py -d mnist
python generate_adv.py -d cifar
python generate_adv.py -d svhn

Running Detectors

Detection methods are being migrated. Currently implemented:

• KDE (Kernel Density Estimation) - utility functions
• LID (Local Intrinsic Dimensionality) - utility functions

Implementation Status

Completed:

• ✅ Basic directory structure
• ✅ Common utilities
• ✅ CNN models for MNIST, CIFAR-10
• ✅ Training script
• ✅ Adversarial generation script
• ✅ KDE utility functions
• ✅ LID utility functions

In Progress:

• 🔄 Detection scripts (detect_kde.py, detect_lid.py, etc.)
• 🔄 MagNet detection method
• 🔄 Other detection methods (FS, NSS, NIC, multiLID)

TODO:

• ❌ SVHN CNN model (requires ResNet18 implementation)
• ❌ ImageNet support
• ❌ Complete detector implementations
• ❌ Testing and validation

Notes

• Models are saved in HDF5 format (.h5) instead of PyTorch format (.pt)
• TensorFlow uses NHWC format by default (no need to transpose for CNNs)
• Some advanced attacks may require additional implementation