Robust Hashing for Image-based Malware Classiﬁcation

Wei-Chung Huang, Fabio Di Troia and Mark Stamp

Department of Computer Science, San Jose State University, San Jose, California, U.S.A.

Keywords:

Robust Hashing, Malware, Support Vector Machine, Wavelet, Distributed Source Coding.

Abstract:

In this paper, we compare and contrast support vector machine (SVM) classiﬁers to robust hashing based

strategies for the malware classiﬁcation problem. For both the SVM and robust hashing approaches, we

treat each executable ﬁle as a two-dimensional image. We experiment with two image-based robust hashing

techniques, one that relies on wavelet analysis, and one that uses distributed coding. For our support vector

machine experiments, we consider an image-based feature that deals with horizontal edges. While the SVM

performs slightly better, there are some potential advantages to robust hashing for malware detection.

1 INTRODUCTION

Malware is software that is intentionally designed to

cause harm to computer systems (Aycock, 2006). Due

to our heavy reliance on computers in general, and

software in particular, malware classiﬁcation is a vi-

tally important topic in information security. Sig-

nature scanning (i.e., pattern matching) is the most

common form of malware detection, and hence mal-

ware writers have developed many concealment stra-

tegies aimed at defeating standard signature scanning

techniques (Aycock, 2006). These concealment stra-

tegies result in malware families consisting of large

numbers of related variants. Hence, any practical mal-

ware detection or classiﬁcation strategy must be ai-

med at entire families, rather than individual malware

samples.

Malware classiﬁcation can be based on static ana-

lysis or dynamic analysis, or a combination of the

two. Static malware analysis relies on features that

can be extracted without executing (or emulating) the

code—mnemonic opcodes are a well-known example

of a static feature (Yewale and Singh, 2016; ?). Dyn-

amic malware analysis considers the behavior of soft-

ware, which requires the software to be executed (or

emulated) when extracting such features. API calls

are a common dynamic feature (Alazab et al., 2010).

In this research, we treat malware samples as

two-dimensional images and extract static featu-

res. We then apply techniques from image robust

hashing (Venkatesan et al., 2000) and machine lear-

ning to analyze the samples based on these features.

Our goal is to compare the effectiveness of a relatively

straightforward and intuitive machine learning techni-

que to robust hashing based strategies. The authors

are not aware of any previous research that considers

robust hashing techniques in the context of malware

detection or analysis.

The remainder of this paper is organized as fol-

lows. In Section 2 we discuss relevant background

topics, including treating malware as images, an in-

troduction to robust hashing, the features we consider,

and so on. In Section 3, we give our experimental re-

sults, and provide context for these results. Finally,

Section 4 contains our conclusion and a brief discus-

sion of possible future work.

2 BACKGROUND

In this section, we ﬁrst motivate the work in this pa-

per, by presenting examples of executable ﬁles vie-

wed as images. Next, we give a brief overview of

related previous work, followed by an introduction to

robust hashing. We conclude this section with a dis-

cussion of the various image features we consider, and

an outline of the machine learning technique that we

use.

2.1 Executables as Images

We can treat any binary ﬁle as an image by simply in-

terpreting the data as a two-dimensional array of va-

lues and converting to the desired image format. In-

tuitively, we might expect that some common mal-

ware obfuscation techniques can be mitigated by vie-

Huang, W-C., Troia, F. and Stamp, M.

Robust Hashing for Image-based Malware Classiﬁcation.

DOI: 10.5220/0006942204510459

In Proceedings of the 15th International Joint Conference on e-Business and Telecommunications (ICETE 2018) - Volume 1: DCNET, ICE-B, OPTICS, SIGMAP and WINSYS, pages 451-459

ISBN: 978-989-758-319-3

451

wing an executable as an image. That this is indeed

the case is strikingly illustrated in Figure 1, which in-

cludes images corresponding to three samples from

each of three different malware families. For com-

parison, benign executables (as images) appear in Fi-

gure 2.

Figure 1: Images from malware families Agent.FYI (3 left-

most), C2LOP.P (3 top right), and Alueron.gen!J (3 bottom

right).

Figure 2: Examples of benign executables as images.

From the examples in Figures 1 and 2, it appears

that image-based analysis has the potential to be a

strong technique for discriminating between malware

families, and possibly also for discriminating between

malware and benign applications. Furthermore, we

can easily deduce the structure of an executable from

its image. For example, in Figure 3, we can clearly

see the various segments that are present in this exe-

cutable.

2.2 Previous Work

The paper (Nataraj et al., 2011) considers malware

classiﬁcation based on image analysis, and it is shown

that malware samples from the same family tend to

have similar image characteristics. In (Torralba et al.,

2003), high level image features known as GIST des-

criptors are applied to the malware problem, and

strong results are obtained. The work in (Yajamanam

et al., 2018), builds on previous image-based malware

work and obtains improved results using neural net-

works and transfer learning.

Figure 3: Code sections corresponding to a malware image.

The paper (Venkatesan et al., 2000) proposed ro-

bust image hashing to handle the proliferation of di-

gital images. This robust hashing technique has many

uses, including image indexing and image authenti-

cation (Lin and Chang, 1998; Schneider and Chang,

1996; Zhao et al., 2013). The robust image hashing

method considered in (Zhao et al., 2013) uses both

global and local features, while (Victor, 1994) pre-

sents a framework for the study of texture perception.

Feature extraction for robust image hashing is

considered in (Monga et al., 2006), where the propo-

sed process consists of feature vector extraction fol-

lowed by compression. The technique in (Venkate-

san et al., 2000) relies on wavelet decomposition for

image feature extraction.

Another example of a novel robust image hashing

technique can be found in the paper (Johnson and

Ramchandran, 2003), where dithering and distributed

source coding form the based for the technique. In

this case, the syndrome serves as the hash value.

2.3 Robust Hashing

There are many types of hashing. For example, cryp-

tographic hash functions have a wide range of uses in

security-related applications, ranging from digital sig-

natures to blockchain technology (Nakamoto, 2009).

A crucial property of a cryptographic hash is colli-

sion resistance, in the sense that it is computationally

infeasible to ﬁnd distinct inputs that hash to the same

value (Stamp, 2017).

Robust hashing is “robust” in the sense that simi-

lar objects are supposed to hash to the same (or si-

milar) value. Note that this is essentially the oppo-

site of the collision resistance property of a crypto-

graphic hash. Robust image hashing has proven use-

BASS 2018 - International Workshop on Behavioral Analysis for System Security

452

ful in biometric authentication (Sutcu et al., 2005),

image watermark veriﬁcation (Schneider and Chang,

1996), and image indexing (Venkatesan et al., 2000),

for example.

Again, the goal of robust image hashing is to iden-

tify similar images, that is, similar images should hash

to the same value. As proposed by (Monga et al.,

2006), from a high level, the process consists of ex-

tracting relevant features, generating an intermediate

“hash” value, then applying a compression step to ge-

nerate the ﬁnal hash value, as summarized in Figure 4.

Note that the purpose of the compression step is to ne-

gate minor differences or remove noise so that simi-

lar intermediate hash values are clustered to the same

group.

Figure 4: Process of robust hashing.

One approach that has been proposed for the com-

pression step is the use of error-correcting codes (Ven-

katesan et al., 2000). Such codes have the desirable

effect of ignoring small perturbations. A related ap-

proach can be found in (Johnson and Ramchandran,

2003), where a distributed source coding is consi-

dered. In this paper, we experiment with both of

these techniques for the compression step. For the

error-correcting code case, we use a simple Hamming

code (Hamming, 1950).

Hamming codes are a class of linear error-

correcting codes. Here, we use the Hamming(7, 4)

code, a popular conﬁguration with 4 data bits and 3

parity bits, which enables us to correct any 1 bit error.

Distributed source coding (DSC) reduces the com-

putational burden of encoding (Pradhan and Ram-

chandran, 2003). Here, we employ a Wyner-Ziv

encoder as developed and discussed in (Wyner and

Ziv, 1976). A detailed discussion of DSC is beyond

the scope of this paper; for additional information,

see (Johnson and Ramchandran, 2003), for example.

2.4 Image Features

Image features can be categorized as either local or

global. In this paper, we employ two local features

and six global features. The local features we consi-

der are the following.

Local Binary Pattern (LBP) is a well-known tex-

ture descriptor that is widely used in facial recog-

nition (Ahonen et al., 2006). The LBP is found by

extracting the local pixel contrast, relative to each

pixel. The resulting histogram serves as the LBP

feature vector.

Histogram of Oriented Gradient (HOG) is another

texture descriptor that is wildly used due to its ro-

bustness (Dalal and Triggs, 2005). Whereas LBP

is based on contrast, HOG is based on the local

gradient.

In contrast to local features, global feature are fo-

cused on the entire image instead of local regions.

The global features we consider are the following.

Horizontal Edges is a feature that has been used in

many digital image processing applications. Ba-

sed on the executable images in Figure 1, for ex-

ample, it appears that a horizontal edge feature

will be relevant for the malware classiﬁcation pro-

blem considered here.

Speciﬁcally, we extract the distribution of hori-

zontal edges from the edge magnitudes. Figure 5

illustrated this process. Note that we apply a low-

pass ﬁlter, which serves to enhance the edges.

We then project the edge map magnitudes onto

a one-dimensional space. Figure 6 illustrates the

projection we obtain from the graph in Figure 7.

We give a side-by-side comparison of the image

and its horizontal edge projection in Figure 6.

Pixel Intensity is a useful global feature. This fea-

ture consists simply of a histogram of pixel inten-

sities over the entire image.

Contrast is one of the most widely used features.

The contrast tells us the variability in the bright-

ness over an image.

Median Filter consists of low-pass ﬁlters with vari-

ous noise reduction techniques applied. A median

ﬁlter is designed to preserve edges while smoo-

thing the image and hence this feature appears to

be a good candidate for the malware image analy-

sis problem.

Frequency Distribution is based on quantized dis-

crete cosine transformation (DCT) coefﬁcients.

This feature measures the overall variability in the

frequency domain.

Wavelet Transform can be used to generate featu-

res based on a series of different image resoluti-

ons (Daubechies, 1990). Roughly speaking, wa-

velet transforms enable us to analyze the signal

in the frequency domain, while retaining tempo-

ral information. For our purposes, we decompose

the image to several sub-bands as illustrated in Fi-

gure 8. We then compute statistical features from

each band. For example, we compute the mean

pixel value on the coarse band and the variance

on the ﬁne band.

Robust Hashing for Image-based Malware Classiﬁcation

453

Figure 5: Horizontal edge feature extraction.

Figure 6: Horizontal edge magnitude feature vector.

2.5 Support Vector Machines

We compare our robust hashing technique to a ma-

chine learning classiﬁer. For our machine learning ex-

periments, we have chosen to use a support vector ma-

chine (SVM) (Cristianini and Shawe-Taylor, 2000).

SVMs are well-known and widely used in many dom-

ains.

An SVM classiﬁer attempts to determine a se-

parating hyperplane where the margin (i.e., the mi-

nimum distance from the separating hyperplane to

either class) is maximized. However, the data itself

need not be linearly separable, in which case no se-

parating hyperplane exists in the input space. In the

SVM technique, we can shift the problem to a hig-

her dimensional feature space by using a non-linear

transformation via the so-called kernel trick. By mo-

ving to a higher dimension, there is a much greater

chance that the training data will be linearly separa-

ble. The real beauty of SVMs is that we pay virtually

no computational penalty for working in this higher

dimensional feature space.

3 EXPERIMENTS

In this section, we give our experimental results. We

apply three versions of robust hashing to the malware

classiﬁcation problem. For comparison, we also ap-

ply an SVM to the same malware classiﬁcation pro-

blem. But ﬁrst, we discuss our dataset.

3.1 Dataset

For the experiments in this paper, we use the Malimg

dataset (Nataraj et al., 2011). As can be seen in Ta-

ble 1, this dataset consists of 9,342 grayscale images

representing binaries from 25 malware families of a

variety of different types. Figure 9 gives examples of

images found in this dataset.

Here, we deal with the problem of classifying the

Malimg images into their 25 families. First, we con-

sider an SVM, then we apply robust hashing techni-

ques.

3.1.1 SVM for Classiﬁcation

We extracted the features discussed in Section 2.4 for

each of the malware images in the Malimg dataset.

For all experiments discussed here, we use ﬁve-fold

cross-validation, that is, we randomly partition the

samples into ﬁve equal sized sets, S

, S

In the ﬁrst “fold,” we train on sets S

, S

, re-

serving S

for testing; in the second fold, we train

on S

, S

, reserving S

for testing, and so on.

Cross validation serves to minimize any bias in the

data, while also maximizing the number of test cases.

Figure 10 gives our results for the classiﬁcation

problem using an SVM based on the edge feature,

as discussed in Section 2.4. In this case, we obtain

an overall classiﬁcation accuracy of about 84% if we

consider the per family average. On the other hand,

the classiﬁcation accuracy is over 93% on a per sam-

ple basis, indicating that some of the less numerous

BASS 2018 - International Workshop on Behavioral Analysis for System Security

454

Figure 7: Horizontal edge feature vector and image.

Figure 8: Sub-bands of wavelet decomposed image.

Figure 9: Malimg examples.

families are among those that are most difﬁcult to dis-

tinguish. Furthermore, these results show that only a

relatively small number of families are not classiﬁed

Table 1: Malimg Dataset.

Family Type Samples

1 Adialer.C Dialer 125

2 Agent.FYI Backdoor 116

3 Allaple.A Worm 2949

4 Allaple.L Worm 1591

5 Alueron.gen!J Trojan 198

6 Autorun.K Worm 106

7 C2LOP.gen!G Trojan 200

8 C2LOP.P Trojan 146

9 Dialplatform.B Dialer 177

10 Dontovo.A Trojan DL 162

11 Fakerean Rogue 381

12 Instantaccess Dialer 431

13 Lolyda.AA 1 PWS 213

14 Lolyda.AA 2 PWS 184

15 Lolyda.AA 3 PWS 123

16 Lolyda.AT PWS 159

17 Malex.gen!J Trojan 136

18 Obfuscator.AD Trojan DL 142

19 Rbot!gen Backdoor 158

20 Skintrim.N Trojan 80

21 Swizzor.gen!E Trojan DL 128

22 Swizzor.gen!I Trojan DL 132

23 VB.AT Worm 408

24 Wintrim.BX Trojan DL 97

25 Yuner.A Worm 800

with high accuracy. For example, if we remove the

seven families with the lowest accuracy from the set

of 25 families, then we obtain an accuracy in excess

of 98%. Overall, these results show that a simple and

intuitive horizontal edge feature is extremely useful

for this image-based malware classiﬁcation problem.

Having determined appropriate parameters for the

features, we now consider the problem of feature ana-

lysis. Speciﬁcally, we consider the tradeoff between

the features used and accuracy. We apply both univa-

riate feature selection (UFS) and recursive feature eli-

mination (RFE). UFS simply consists of testing each

feature independently and ranking them based on the

Robust Hashing for Image-based Malware Classiﬁcation

455

Figure 10: Classiﬁcation accuracy using SVM.

accuracy of the resulting models. In RFE, we remove

the lowest performing feature, based on linear SVM

weights, then retrain the SVM on the reduced feature

set. We generally expect RFE to be somewhat more

reliable, as it accounts for the interactions between fe-

atures.

Figure 11 depicts our UFS results for our global

image features in the form of a bar graph. Perhaps not

surprisingly, the horizontal edge feature outperforms

all other features.

Figure 11: UFS results based on accuracy.

We further analyze the horizontal edge feature

using RFE. Recall that this edge feature is 256 dimen-

sional. When we apply RFE to this speciﬁc feature,

we obtain the results in Figure 12. In this case, we

ﬁnd that by using 169 of the 256 dimensions, we can

obtain optimal accuracy. We use this 169 dimensional

set of horizontal edge features in the subsequent SVM

experiments.

Using an SVM and the reduced horizontal edge

feature as discussed above, we have achieved a clas-

siﬁcation accuracy of 92% for the 25 families in the

Malimg dataset. Figure 13 gives confusion matrix (in

the form of a heatmap) for this experiment. Note that

in this ﬁgure, the families are numbered as in Table 1.

Figure 12: RFE result on horizontal edge feature.

Family 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25

Figure 13: Confusion matrix for SVM as a heatmap.

Next, we apply two different robust hashing ap-

proaches to the malware classiﬁcation problem. Re-

call that the robust hashing problem consists of fe-

ature vector extraction, followed by a compression

step. In this research, we implement and test the

image robust hashing technique from (Venkatesan

et al., 2000), and the technique from (Johnson and

Ramchandran, 2003), as well as a combined model.

3.1.2 Robust Hashing based on Wavelets

The robust hashing technique that is given in (Ven-

katesan et al., 2000) uses wavelet based analysis and

an error-correcting decoder. In our implementation,

we extract statistical information from different fre-

quency bands and group together similar feature vec-

tors based on an error-correcting decoder. Speciﬁ-

cally, we use a 2-dimensional Haar wavelet transform

and a ﬁve level decomposition. We then use a scalar

quantization to normalize the feature vector over 128

levels. After all of these wavelet based features are

acquired, we use a Hamming(7, 4) decoder to “cor-

rect” the sequence of quantized 7-bits values, thus ge-

nerating a vector that represents an image family.

Classiﬁcation results using this approach are sum-

marized in Figure 14, where we obtain an overall

accuracy of 79%. Although this is signiﬁcantly lower

than the 92% accuracy we obtained with an SVM, if

BASS 2018 - International Workshop on Behavioral Analysis for System Security

456

we limit the analysis to the top 19 of the 25 families,

we have a comparable accuracy of 90%, while with 11

families we can attain well over 95% accuracy.

Figure 14: Classiﬁcation result for wavelet approach.

3.1.3 Robust Hashing based on Distributed

Coding

Next, following the general approach considered in

the paper (Johnson and Ramchandran, 2003), we

apply a distributed coding technique to the robust

hashing problem. As implemented in the rese-

arch (Varodayan et al., 2007), we have employed a

Wyner-Ziv encoder based on DCT compression, and

we use a Hamming code to produce a syndrome. The

syndrome serves as the ﬁnal robust hash value.

For this approach, we achieve slightly over 83%

classiﬁcation accuracy for all 25 families. Recall that

the wavelet based approach was only able to attain

about 79% accuracy. This is a signiﬁcant impro-

vement, but still not on par with the SVM results

which achieved an accuracy of more than 92%.

Figure 15: Classiﬁcation result for distributed coding ap-

proach.

3.1.4 A Mulitphase Approach to Robust

Hashing

From the robust hashing classiﬁcation results above,

we observe that for most families, one approach is

more accurate than the other. For example, the family

Autorun.k is classiﬁed well by the wavelet based clas-

siﬁer, but poorly by the distributed coding based clas-

siﬁer. Here, we combine both our wavelet based and

our distributed coding based robust hashes—we refer

to the resulting combination as a multiphase robust

hash.

Intuitively, we can combine our two robust

hashing techniques by scoring with both techniques,

and in cases where the classiﬁcations disagree, we

select the classiﬁcation with the higher accuracy for

its model. Figure 16 gives our results for this mul-

tiphase approach. Note that we achieve an accuracy

above 87% over all 25 families, and we can attain an

accuracy above 90% by discarding only one family.

While this is not quite as strong as the 92% accuracy

that we achieved with an SVM, it is comparable.

Figure 16: Classiﬁcation result for multiphase approach.

Finally, Figure 17 compares the classiﬁcation

accuracy for the three robust hashing techniques con-

sidered. As with all of our classiﬁcation experiments,

these results are based on the 25 families in the Ma-

limg dataset.

4 CONCLUSION

In this paper, we considered the problem of classi-

fying images corresponding to 25 malware families.

We compared a straightforward SVM classiﬁer based

on a simple and intuitive horizontal edge feature to

robust hashing techniques. For the SVM, we care-

fully considered feature analysis issues, while for the

robust hashing case, we experimented with three ap-

proaches. Overall, an optimized SVM performed so-

Robust Hashing for Image-based Malware Classiﬁcation

457

Figure 17: Comparison of robust hashing approaches.

mewhat better, but our best robust hashing results are

comparable. With respect to robust hashing, we ﬁnd

that these techniques perform better than the SVM on

most families, but a few malware families are extre-

mely difﬁcult to classify. In contrast, the SVM yields

more consistent results over the 25 families.

In spite of the slightly lower accuracy, there are

some potential advantages to robust hashing. Per-

haps the biggest of these advantages is that in ro-

bust hashing, it is easy to add new families as they

appear—we simple generate a syndrome (which, in

effect, deﬁnes a cluster) for the new family and the re-

mainder of the robust hashing classiﬁcation model is

unchanged. In contrast, for an SVM, we would need

to retrain the model each time a new family is added.

For large problems, the SVM training cost—in terms

of both time and computational resources—would be

substantial.

It appears that malware analysis is a novel appli-

cation of robust hashing. And it is worth noting that

robust hashing is a fairly general and somewhat amor-

phous concept. Hence, a multitude of variations on

robust hashing can be considered. In addition to tes-

ting some of the many different possible forms or ro-

bust hashing, future work could include an analysis of

additional features. We could also obtain a more ﬁne-

grained view by analyzing each section of an executa-

ble (i.e., .text, .rdata, and so on) separately. In ad-

dition, hybrid classiﬁcation techniques involving ro-

bust hashing and any of a variety of machine learning

techniques would be an interesting topic for further

research. For example, we could apply various ro-

bust hashing techniques to a variety of image features,

then apply a machine learning classiﬁer to the results

of these robust hashing algorithms.

Within the robust hashing paradigm, it would be

worthwhile to experiment with more sophisticated co-

ding techniques, such as Reed-Muller codes or trellis-

coded modulation (TCM). Another type of hybrid

model would consist of using non-coding based com-

pression strategies within a robust hashing scheme—

techniques such as K-means, EM clustering, k-nearest

neighbor, and Gaussian mixture models, for exam-

ple, would ﬁt naturally within a robust hashing fra-

mework.

REFERENCES

Ahonen, T., Hadid, A., and Pietikainen, M. (2006). Face

description with local binary patterns: Application to

face recognition. IEEE Transactions on Pattern Ana-

lysis and Machine Intelligence, 28(12):2037–2041.

Alazab, M., Venkataraman, S., and Watters, P. (2010). To-

wards understanding malware behaviour by the ex-

traction of api calls. In Proceedings of the 2010 Se-

cond Cybercrime and Trustworthy Computing Works-

hop, CTC ’10, pages 52–59. IEEE Computer Society.

Aycock, J. (2006). Computer Viruses and Malware. Sprin-

ger.

Cristianini, N. and Shawe-Taylor, J. (2000). An Introduction

to Support Vector Machines and Other Kernel-Based

Learning Methods. Cambridge University Press.

Dalal, N. and Triggs, B. (2005). Histograms of oriented gra-

dients for human detection. In IEEE Computer Society

Conference on Computer Vision and Pattern Recogni-

tion, CVPR 2005, pages 886–893. IEEE.

Daubechies, I. (1990). The wavelet transform, time-

frequency localization and signal analysis. IEEE

Transactions on Information Theory, 36(5):961–1005.

Hamming, R. W. (1950). Error detecting and error

correcting codes. Bell Labs Technical Journal,

29(2):147–160.

Johnson, M. and Ramchandran, K. (2003). Dither-based se-

cure image hashing using distributed coding. In Pro-

ceedings of 2003 International Conference on Image

Processing, ICIP 2003, pages 751–754. IEEE.

Lin, C.-Y. and Chang, S.-F. (1998). Generating robust digi-

tal signature for image/video authentication. In Mul-

timedia and Security Workshop at ACM Multimedia,

pages 49–54.

Monga, V., Banerjee, A., and Evans, B. L. (2006). A clus-

tering based approach to perceptual image hashing.

IEEE Transactions on Information Forensics and Se-

curity, 1(1):68–79.

Nakamoto, S. (2009). Bitcoin: A peer-to-peer electronic

cash system. https://bitcoin.org/bitcoin.pdf.

Nataraj, L., Karthikeyan, S., Jacob, G., and Manjunath, B.

(2011). Malware images: Visualization and automatic

classiﬁcation. In Proceedings of the 8th International

Symposium on Visualization for Cyber Security, Viz-

Sec’11, pages 4:1–4:7. ACM.

Pradhan, S. S. and Ramchandran, K. (2003). Distributed

source coding using syndromes (DISCUS): Design

and construction. IEEE Transactions on Information

Theory, 49(3):626–643.

BASS 2018 - International Workshop on Behavioral Analysis for System Security

458

Santos, I., Brezo, F., Nieves, J., Penya, Y. K., Sanz, B., La-

orden, C., and Bringas, P. G. (2010). Idea: Opcode-

sequence-based malware detection. In International

Symposium on Engineering Secure Software and Sys-

tems, pages 35–43. Springer.

Schneider, M. and Chang, S.-F. (1996). A robust content ba-

sed digital signature for image authentication. In Pro-

ceedings of 1996 International Conference on Image

Processing, ICIP 1996, pages 227–230. IEEE.

Stamp, M. (2017). Introduction to Machine Learning with

Applications in Information Security. Chapman and

Hall/CRC, Boca Raton.

Sutcu, Y., Sencar, H. T., and Memon, N. (2005). A se-

cure biometric authentication scheme based on robust

hashing. In Proceedings of the 7th Workshop on Mul-

timedia and Security, pages 111–116. ACM.

Torralba, A., Murphy, K. P., Freeman, W. T., and Rubin,

M. A. (2003). Context-based vision system for place

and object recognition. https://www.cs.ubc.ca/ murp-

hyk/Papers/iccv03.pdf.

Varodayan, D., Lin, Y.-C., Mavlankar, A., Flierl, M., and

Girod, B. (2007). Wyner-ziv coding of stereo images

with unsupervised learning of disparity. In Procee-

dings of Picture Coding Symposium.

Venkatesan, R., Koon, S.-M., Jakubowski, M. H., and Mou-

lin, P. (2000). Robust image hashing. In Proceedings

of 2000 International Conference on Image Proces-

sing, volume 3 of ICIP 2000, pages 664–666. IEEE.

Victor, J. D. (1994). Images, statistics, and textures: Impli-

cations of triple correlation uniqueness for texture sta-

tistics and the julesz conjecture: Comment. Journal of

the Optical Society of America A, 11(5):1680–1684.

Wyner, A. and Ziv, J. (1976). The rate-distortion function

for source coding with side information at the decoder.

IEEE Transactions on Information Theory, 22(1):1–

10.

Yajamanam, S., Selvin, V. R. S., Troia, F. D., and Stamp,

M. (2018). Deep learning versus gist descriptors for

image-based malware classiﬁcation. In Proceedings

of the 2nd International Workshop on Formal Methods

for Security Engineering, ForSE 2018.

Yewale, A. and Singh, M. (2016). Malware detection based

on opcode frequency. In Advanced Communication

Control and Computing Technologies (ICACCCT),

2016 International Conference on, pages 646–649.

IEEE.

Zhao, Y., Wang, S., Zhang, X., and Yao, H. (2013). Ro-

bust hashing for image authentication using zernike

moments and local features. IEEE Transactions on

Information Forensics and Security, 8(1):55–63.

Robust Hashing for Image-based Malware Classiﬁcation

459