Adversarial Attack to Capsule Networks

This repo. is to research adversarial attack performance for CapsNets.

In those days, deep learning has shown an attractive results on several applications (speech / image recognization). However, from 2014, there are research papers that deep learning can be easiy fool by very low noise. In NIPS 2017 workshop, Ian Goodfellow opened adversarial non-targeted/targeted attack and defense competition for this problem, and In ICLR2018, lots of researches for adversarial examples are submitted. In this situation, we read G. Hinton's paper, Dynamic routing between capsules, and tried to validate the robustness of CapsNets on the paper to well-known adversarial attack.

The CapsNets part on this repo. is based on InnerPeace-Wu's ones, https://github.com/InnerPeace-Wu/CapsNet-tensorflow.

Settings

This implementation is on tensorflow 1.2.1, and the detailed setting is followed InnerPeace-Wu's one.

Attackers

Fast Gradient Sign Method (FGSM)

Fast Gradient Sign Method is to generate adversarial images with gradient information from loss function and true label.

https://arxiv.org/pdf/1607.02533.pdf

python attack_gsm.py --max_iter=1

This module is based on gongzhitaao's git and sample attack codes from Google Brain.

https://github.com/gongzhitaao/tensorflow-adversarial

Basic Iteration (basic iter.)

Basic Iteration is to generate adversarial images with iteratively running FGSM.

https://arxiv.org/pdf/1607.02533.pdf

python attack_gsm.py --max_iter=<iter_num>

This module is based on gongzhitaao's git and sample attack codes from Google Brain.

https://github.com/gongzhitaao/tensorflow-adversarial

Least-likely Class Method (step l.l)

Least-likely Class Method is to generate adversarial images by not using true label but "least-likely" label.

https://arxiv.org/pdf/1607.02533.pdf

python attack_llcm.py --max_iter=1

This module is based on gongzhitaao's git and sample attack codes from Google Brain.

https://github.com/gongzhitaao/tensorflow-adversarial

Iterative Least-likely Class Method (iter. l.l)

Iterative Least-likely Class Method is to generate adversarial images with iteratively running step l.l.

https://arxiv.org/pdf/1607.02533.pdf

python attack_llcm.py --max_iter=<iter_num>

This module is based on gongzhitaao's git and sample attack codes from Google Brain.

https://github.com/gongzhitaao/tensorflow-adversarial

Adversarial Attack based on adversarial generative learning(AGL)

This module is to generate adversairal noise can make the model mis-predict data added with that.

In here it, three loss terms are as bellowed.

1. reverse cross-entropy for true label (rce) => getting insight from FGSM
2. cross-entropy for the lease label (ce_ll) => getting insight from step l.l
3. cross-entropy for top-1 selected label exept true label (ce_s) => new things

python attack_gan.py

The Agents

Naiive conv model (Baseline)

This module is not consisted with adversarial prior to defense.

The structure of this is as bellowed.

3 convolutional layers with Relu activation function, the kernel and feature size of which are 3 and [64,128,256], respectively. After 3 conv. layers, 2 linear layers with Relu and softmax activation function are used to get prediction of them.

The accuracy for MNIST original test data is about 99%.

Adversarial Training with adversarial examples from FGSM and Basic iter.

This module is to regularize baseline model with FGSM and Basic iter.

On each epoch, adversarial examples are repeatly generated and learned with the agent.

Adversarial Training with adversarial examples from step l.l and iter. l.l

This module is to regularize baseline model with step l.l.

On each epoch, adversarial examples are repeatly generated and learned with the agent.

Experiments Results (MNIST)

• CapsNets is learned with 1.5K iteration, and training/test acc. are 0.995/0.991, respectively.

• Epsilon is calculated with infinity norm

• Max epoch is 50

• Max iter. for basic iter. and iter. l.l is 5.

• gif file is adversarial examples with max_epsilon from 5 to 50

FGSM

Basic iter.

step l.l

iter l.l

Discussion

CapsNet also fall in the trap of every type adversarial examples. This structure can be one of the hint to solve adversarial problem, however experiment results show CapsNets is not free to adversarial attack.