TIAFuzz

A Reusable sparse adversarial test case generation method for CV(computer Vision) software

Code release and supplementary materials for：

"TIAFuzz：Transferable Fuzzing via Distillation for Image-Based Deep Learning Systems"

Datasets

• CIFAR-10
• CIFAR-100
• Flower102

Dependencies

The code was tested with:

• python 3.7.0
• h5py 3.8.0
• ipykernel 6.19.2
• matplotlib 3.7.2
• numpy 1.25.2
• pandas 1.5.3
• scikit-image 0.21.0
• scipy 1.9.3
• torch 1.11.0
• torchvision 0.12.0
• tqdm 4.64.1

Training

Training the surrogate model.

   python main.py

Evaluations

Test case generation by the test case generation methods on CIFAR-10 using or without using the surrogate model trained by SDST

python attack.py - method_name_index 7

Index	Method	Surrogate Model	Index	Method	Surrogate Model
0	GreedyFool	ResNet18	1	GreedyFool	EMA-ResNet18
2	PGD	ResNet18	3	PGD	EMA-ResNet18
4	SparseFool	ResNet18	5	SparseFool	EMA-ResNet18
6	GMI	ResNet18	7	GMI	EMA-ResNet18

Ablations

1. Training the surrogate model without using EMA

   python main_ablation.py  #train the surrogate model
   cd Ablation_EMA
   python attack.py
   

2. Attack the surrogate model by GMI without using momentum

    cd Ablation_momentum
    python attack.py