Malware Image Classfication

This repository covers the malware image classification and malware image generation. We use Microsoft Malware Classification Challenge (BIG2015) dataset. To classify the malware examples, we utilize the image classification task using DNN models, e.g., InceptionV3. Given the binary samples, we can convert them into the images based on the Incremental Coordinate(IC) Method. Readers may refer to this paper to generate the malware images. Malware image samples generated by IC method are prsented as below. We depict two samples per class from 1 to 9.

Preprocessing

Dataset

Microsoft Malware Classification Challenge (BIG2015)

Family name	Number of samples	Type
Ramnit	1,541	Worm
Lolipop	2,478	Adware
Kelihos ver3	2,942	Backdoor
Vundo	475	Trojan
Simda	42	Backdoor
Tracur	751	Trojan Downloader
Kelihos ver1	398	Backdoor
Obfuscator.ACY	1,228	Any kind of obfuscated malware
Gatak	1,013	Backdoor
Sum	10,868

How to convert binary to images

To apply DNN-based malware classification, each malware sample should be converted into a corresponding image of fixed size.
This figure illustrates how malware files are converted into Incremental Coordinate image files. First, the malware binary is read in 4-byte increments. The first 2 bytes are converted to x coordinates, the last 2 bytes are converted to y coordinates, the value represented by those coordinates is incremented by 1. This process is performed until the end of the file is reached. This generates a 256 by 256 gray scale image from each malware sample. The key advantage of visualizing a malicious PE file is that it can perform malware detection/classification without running malware. In addition, small changes that were not detectable in malware binaries may be detectable in malware converted to images

Split the dataset

We divided the dataset into 9:1 ratios using the following code.

cd classification
python split.py # spliting the train and test data

Training the Malware Classifier

> CUDA_VISIBLE_DEVICES=0 python main.py \
    --batch_size 100 \
    --lr 0.01 \
    --name checkpoint

batch_size is a size of batch.
lr is a learning rate to update various losses.
name is a name of folder. checkpoint will be stored on this folder.

Generate AEs with additive gaussian white noise

You can easily generate the noisy samples using below code.

python make_noise.py

Note that, we set the stddev as small enough value since generated samples are supposed to be indistinguishable with the original samples. We empirically found that 0.001 does not hurt structural features yet effective enough to deceive the target classifier. Additive noise will incur the error amplification effects through the intermediate layer of the classifier, resulting in misclassification.

noise = torch.zeros_like(img).cuda()
noise.data.normal_(0, 0.001)
img = noise+img

Evaluate the noise data with pre-trained classifier

Noise inputs will be stored at noise folder, and you can evaluate the pre-trained classifier with the noise inputs. Detail of the experimental results can be found on the original paper, Resilience against Adversarial Examples: Data-Augmentation Exploiting Generative Adversarial Networks, TIIS, 2021

python test.py # output the accuracy

Training cDCGAN

Before training the model, we randomly pick the 32 images per class.

python pick.py # sampling 32 images per class
python train.py # training the cDCGAN model

Re-Training the classifier

Once cDCGAN is fully trained, we can leverage the model to generate extra data to enhance the baseline model. In the simple experiment, we adjust the number of samples and evaluate the performance based on the classification accuracy.

Evaluation

Metric/Test	Baseline	n=10	n=20	n=100
original(balanced)	97.4(96.4)	97.4(95.4)	97.2(95.2)	97.5(95.4)
AEs(balanced)	71.1(59.2)	74.56(66.78)	75.11(70.16)	67.3(65.41)
Average FPR	3.78	2.97	3.0	3.8

Jeffkang-94 / Resilience_against_AEs