Pytorch_CIFAR100&Adversarial Attack

This is my MachineLearing coursework and My work is Based On This Repo

Main Work

• Use ResNet50、MobileNet、VGG19 as Backbone Network Respectively
• Use Tricks such as Xavier Init、LabelSmoothing ON VGG19
• Use FGSM and PGD algorithms to attack VGG19 model

Classification Task Results

network	params	top1 err	top5 err	total epoch
MobileNet	3.3M	32.50	9.67	160
VGG19	39.0M	28.82	10.58	160
ResNet50	23.7M	22.33	5.50	160

• Based On VGG19,I also did some experiments.

Tricks	top1 err	top5 err	total epoch	ACC
Original	28.82	10.58	160	71.18%
WarmUp&XavierInit	27.71	10.49	160	72.29%
No Bias Decay	29.86	11.96	160	drops about 1%
Label Smoothing	26.44	10.92	160	73.56%

Adversairal Attack Results

There is some error in my code when using FGSM algorithm, cuz if you settle eps=0,you can still get a good attack effect(acc from nearly 72% to 26%).Maybe fix in some weeks

eps	top1 err	top5 err	ACC
/	28.82	10.58	71.18%
0	74.37	57.54	25.63%
0.2	90.00	76.71	10.00%
0.3	92.28	81.06	7.72%

PGD results as Following:（I'm not sure whether it really works or not)

eps	top1 err	top5 err	ACC
/	28.82	10.58	71.18%
0	28.82	10.58	71.18%
0.2	99.69	98.21	0.31%

Pytorch-cifar100

practice on cifar100 using pytorch

Requirements

This is my experiment eviroument, pytorch0.4 should also be fine

• python3.5
• pytorch1.0
• tensorflow1.5(optional)
• cuda8.0
• cudnnv5
• tensorboardX1.6(optional)

Usage

1. enter directory

$ cd pytorch-cifar100

2. dataset

I will use cifar100 dataset from torchvision since it's more convenient, but I also kept the sample code for writing your own dataset module in dataset folder, as an example for people don't know how to write it.

3. run tensorbard(optional)

Install tensorboardX (a tensorboard wrapper for pytorch)

$ pip install tensorboardX
$ mkdir runs
Run tensorboard
$ tensorboard --logdir='runs' --port=6006 --host='localhost'

4. train the model

Train all the model on a Tesla P40(22912MB)

Attention:

You need to specify the net you want to train using arg -net if you want to use some training tricks to improve accuracy,maybe you need to modify train.py(I have implemented some training tricks on VGG16.Although the advance is tiny,I guess on other models maybe you can get a better result.)

$ python train.py -net vgg16

5. test the model

Test the model using test.py

$ python test.py -net vgg16 -weights path_to_vgg16_weights_file

6.Attack the model

$ python FGSM.py
$ python PGD.py

Implementated NetWork

• vgg Very Deep Convolutional Networks for Large-Scale Image Recognition
• resnet Deep Residual Learning for Image Recognition
• mobilenet MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications