Data Leakage in Isolated Virtualized Enterprise Computing Systems

Zechariah Wolf

, Eric C. Larson

and Mitchell A. Thornton

Darwin Deason Institute for Cybersecurity at Southern Methodist University, Dallas, TX, U.S.A.

Keywords:

Power Analysis, Side Channel, Virtual Machine, Enterprise Systems.

Abstract:

Previous literature has shown the effectiveness of power analysis as a side channel attack on cryptosystems.

Power analysis is performed using an oscilloscope to measure power consumption information from hardware

utilized during cryptographic algorithms, in order to extract an encryption key. In this paper, we further explore

the potential of power analysis of side channels for leaking information in enterprise computing systems. By

applying the concept of power analysis more broadly to the power consumption of an entire server rack, rather

than individual hardware components, we ﬁnd that basic patterns in system load can be clearly identiﬁed using

signal processing techniques, demonstrating a potential side channel.

1 INTRODUCTION

In the most general sense, a side channel is any means

by which information about the state of a computing

system is leaked unintentionally. From a defensive

perspective, side channels are important to identify

and mitigate in order to reduce the feasibility of side

channel vulnerabilities in otherwise secure algorithms

and systems. An attacker can use side channels to ex-

tract information, and a “side channel attack” is de-

ﬁned as an exploit of a particular side channel to ex-

tract private or sensitive information from a system.

Side channel attacks are particularly dangerous be-

cause the vulnerability exists not in an algorithm or

system itself, but in the nature of how a system physi-

cally works. For example, an otherwise cryptograph-

ically secure encryption algorithm could be vulnera-

ble to a side channel attack because of information

leaked from the hardware components on which the

algorithm was performed (Szefer, 2019).

However, side channels do not only represent a

potential weakness. Recent work has shown that

side channel data can be used defensively to iden-

tify anomalous behavior in a system that could be

indicative of malware. As an example, Taylor et al.

demonstrated a machine learning approach for early

detection of a ransomware attack using the side chan-

nel of physical sensor data as an indicator of anoma-

lous system behavior (Taylor et al., 2021). Trained

https://orcid.org/0000-0002-2226-0327

https://orcid.org/0000-0001-6040-868X

https://orcid.org/0000-0003-3559-9511

models were able to identify subtle, randomly timed

ﬁle encryption operations happening in a simulated

ransomware attack with high accuracy. Similarly,

Khan et al. demonstrated a deep learning approach

for detecting various malware attacks using electro-

magnetic side channel data emanating from an FPGA

(Khan et al., 2019). After being trained to recog-

nize the state of the system under normal operations,

a deep neural network was able to accurately detect

the anomalous conditions caused by simulated mal-

ware attacks, with perfect accuracy in some test cases.

Thus, side channel investigations are important not

only for identiﬁcation and mitigation of vulnerabili-

ties, but also for exploring use cases in defensive strat-

egy.

In this paper, we focus on a power analysis of

a side channel identiﬁed in an enterprise comput-

ing system. However, the data capture we employ

uses a non-traditional methodology that leverages the

electromagnetic interference (EMI) backscatter that

is generated by a computing system while in oper-

ation. This EMI has been used by a number of re-

searchers for identifying electrical devices and appli-

ances (Patel et al., 2007; Gupta et al., 2010), for iden-

tifying changes in electrical device signature that can

be repurposed for interaction techniques (Gupta et al.,

2011; Chen et al., 2013), and has been shown to leak

information regarding television programming (Enev

et al., 2011). Here, we use shifts in the EMI system

ﬁngerprint to indicate the internal state of a comput-

ing system running a virtual machine (VM), showing

that it can be used as a covert side channel.

118

Wolf, Z., Larson, E. and Thornton, M.

Data Leakage in Isolated Virtualized Enterprise Computing Systems.

DOI: 10.5220/0011691900003405

In Proceedings of the 9th International Conference on Information Systems Security and Privacy (ICISSP 2023), pages 118-123

ISBN: 978-989-758-624-8; ISSN: 2184-4356

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

Table 1: Comparison of related side channel investigations.

Approach Advantages Limitations Reference

EM side channel analysis for

malware detection

Non-invasive, noise re-

silient, high detection

accuracy

Experimentation limited

to embedded systems

(Khan et al.,

2019)

Deep-learning-based power

side channel attack on AES

Attack is effective across

different kinds of devices

Data acquisition requires

direct access to micro-

controller power supply

(Golder et al.,

2019)

PLATYPUS - software-

based power analysis

attacks on cryptosystems

Tested channels are accessi-

ble via software interface

Software-based attack is

limited to Intel CPUs

(Lipp et al.,

2021)

Our approach Non-invasive, demonstrates

information leakage in a VM

environment

Resilience to noise not

yet quantiﬁed

EMI and EMI-adjacent side channel analysis is an

ongoing area of research, with applications ranging

from cryptanalysis to malware detection. A few ex-

amples of similar side channel research with varying

applications is presented in Table 1.

2 BACKGROUND

Power analysis is a method of extracting informa-

tion from a side channel – typically power consump-

tion measurements from hardware components. In

a power analysis attack, an oscilloscope is typically

used to gather data. The data can be analyzed to ex-

tract information about the state of the computing sys-

tem. Power analysis research is not new. In 1999,

Kocher et al. demonstrated a power analysis side

channel attack on the DES encryption algorithm. By

collecting traces of electrical current measurements

over time, they were able to extract an encrpytion

key using two different power analysis approaches

(Kocher et al., 1999). Despite years of advance-

ment in technology and heightened awareness of se-

curity, power analysis of side channels remains a dif-

ﬁcult challenge to overcome, and researchers are still

ﬁnding ways to exploit side channels to break cryp-

tosystems. For instance (Lipp et al., 2021) recently

demonstrated a series of power analysis attacks for

extracting encryption keys through power side chan-

nels identiﬁed in Intel CPUs.

However, looking beyond the scope of cryptosys-

tems, power analysis could have applications in side

channels more broadly. Previous research has inves-

tigated power analysis of speciﬁc hardware compo-

nents used in encryption algorithms in order to leak

speciﬁc data being used in the algorithm. We de-

cided to investigate how this kind of approach could

reveal a side channel at the scale of an entire com-

puting system, rather than individual hardware com-

ponents. Speciﬁcally, it would be of interest to iden-

tify a potential side channel in an enterprise comput-

ing system, where tasks are typically run on virtual

machines, with the hypervisor acting as a layer of ab-

straction between the virtual machine and the physical

hardware.

It is also important to note that the EMI generated

by a computing system is related to the power usage

of the system, but can also be inﬂuenced by other sys-

tems on the same circuit or other devices that cause

backscatter at certain frequencies on the voltage spec-

trum. The EMI generated by the computing system

is most easily seen by performing a kHz range fre-

quency analysis on the circuit near where the system

is obtaining power (see Figure 4). Because this EMI

is generated along the entire circuit, this signature can

be obtained by sampling from any nearby power out-

let. Thus, direct access to the computing system is not

required to carry out this exploit.

2.1 Hypothesis

We hypothesize that in an enterprise computing sys-

tem using virtualization software, information about

the state of a virtual machine is leaked through the

side channel of the server’s power consumption, pro-

vided the server’s power supply is not employing

proper side channel mitigation. We propose and per-

form an experiment to modulate the power consump-

tion in an identiﬁable pattern, expecting that if power

metrics are collected over time, power analysis will

reveal the pattern, showing that the side channel is vi-

able and an attacker can ex-ﬁltrate information about

the state of the virtual machine.

Data Leakage in Isolated Virtualized Enterprise Computing Systems

119

3 EXPERIMENTAL DESIGN

In order to evaluate the potential of a power analysis

side channel in an enterprise computing system with

virtualization, we propose an experimental approach

to collect and analyze power consumption metrics

from a server rack.

3.1 Server Rack

The rack used in this experiment contains ten sepa-

rate servers. The power supplies for each server, as

well as the network switches and other rack compo-

nents, are combined into a single 120V AC wall plug,

that powers the entire rack. The servers on the rack

are running a well-known virtual OS as the host op-

erating system. A second laptop is used to connect

to and control the virtual machines on the server over

LAN. The operating system chosen for the virtual ma-

chines is Kubuntu 22, a distribution of Ubuntu using

the KDE Plasma desktop environment.

In order to modulate the power consumption on

the server, an open-source Linux package called

stress-ng is used

. The stress-ng tool is a command

line interface that allows the user precise control over

a machine’s hardware, enabling creation of highly

customizable system loads. stress-ng is an extensive

package with a large variety of options for generat-

ing system stress, but as a proof concept in this ex-

periment, we use stress-ng speciﬁcally for generating

CPU utilization. The goal with using stress-ng is to

generate a pattern of power consumption that is recog-

nizable and distinguishable from background noise.

To accomplish this, a shell script is used to initiate

a stress-ng load for a certain time interval, and then

sleep for an interval. This should create states of high

and low power consumption on the server that we de-

sire to collect as a side channel, ex-ﬁltrating a binary

stream of information. The shell script used in the ex-

periment is summarized by the following pseudocode:

for ((i=0; i < numCycles; i++)) do

sleep 1 #idle for 1 second

stress-ng --cpu 16 -t 1s

done

3.2 Power Analysis Approach

In order to record measurements of the power con-

sumption, the rack is plugged into a surge protector

and our power line interface (PLI) module is plugged

into a nearby power outlet on the same circuit breaker.

This PLI module acts as a high pass ﬁlter, and is a re-

construction of a circuit originally used in an exper-

https://github.com/ColinIanKing/stress-ng

iment to classify the use of home appliances (Gupta

et al., 2010). This PLI allows us to collect high fre-

quency EMI in the range of about 0 to 50 kHz. The

module connects the power line from the surge pro-

tector to the oscilloscope. The oscilloscope used in

this experiment is from the Picoscope 4000a series.

The oscilloscope is connected via USB to a laptop

which has the required software and drivers to mon-

itor and collect data from the device. Picoscope pro-

vides a monitoring software to show in real-time the

measurements from the connected oscilloscope. We

also make use of a Python wrapper library for the

Picoscope software development kit, which allows

us to collect data from the oscilloscope with a cus-

tom Python script. This script is based on examples

provided by the maintainers of the Python package

. While the stress-ng loads are running on the vir-

tual machine, the laptop connected to the oscilloscope

runs a data collection script, which collects voltage

measurements at a speciﬁed sampling rate, and out-

puts the result to a ﬁle.

Figure 1: Block diagram showing data collection process.

4 EXPERIMENTAL RESULTS

4.1 Initial Exploration of the Hypothesis

Before performing formal experiments, we used Pi-

coscope’s software to observe visualizations of the

data obtained from the oscilloscope in real-time. The

software generates a live frequency spectrum, which

shows voltage gain vs. frequency. We observed this

graph while generating a power consumption square

wave on one of the servers in the rack, the goal being

to see if the modulation in power consumption was

visually discernible. The results were quite clear.

Figure 2 shows the real-time graph when the vir-

tual machine is sitting at idle, with no high-load tasks

running. There is a noticeable spike in the spectrum

at around 30-35 kHz.

Figure 3 shows the same graph when a 100% CPU

utilization stress test is running. The spike that was

present under idle conditions falls dramatically. Dur-

ing the real-time experiment, this fall and rise pattern

correlated exactly with the initiation and ending of the

stress test. Since this rise and fall is occurring at the

https://github.com/picotech/picosdk-python-wrappers

ICISSP 2023 - 9th International Conference on Information Systems Security and Privacy

120

Figure 2: Real-time power spectral density (dBu) vs. fre-

quency (kHz) graph with the virtual machine at idle.

Figure 3: Real-time power spectral density (dBu) vs. fre-

quency (kHz) graph with the virtual machine at 100% CPU

utilization.

30-35 kHz range, this indicates that the data collec-

tion script should sample above 70 kHz in order to

capture this EMI signal properly.

4.2 Signal Processing

The voltage gain vs. frequency graph was useful for

observing the change in real-time, but in order to eval-

uate the possibility of a covert side channel, we must

be able to reliably detect the high and low states oc-

curring over a time range. A transformation that fa-

cilitates this detection is the voltage magnitude spec-

trogram. A magnitude spectrogram can be visualized

like a heat map showing the voltage gain as the inten-

sity and the frequency (along the vertical axis) at each

point in time (along the horizontal axis).

The output of the data collection script is a single

array of values, which represent the voltage signal in

the time domain. In order to produce a spectrogram,

the raw voltage information collected from the os-

cilloscope is processed in a short-time Fourier trans-

form, which produces a complex-valued matrix, such

that each vertical slice represents the frequency spec-

trum of the signal at a given point in time. The matrix

is then decibel-scaled based on the magnitude of the

complex values, in order to reduce the dynamic range

of the output. The ﬁnal result shows the decibel-

scaled voltage intensity at each frequency (vertical

axis), at each point in time (horizontal axis). In par-

ticular, we use a spectrogram with a window length

of 2048 points (about 20 ms), 75% overlap, and von

Hann windowing. The signal processing procedure

is performed using open-source packages in Python.

For signal processing algorithms and spectrogram vi-

sualizations we use Librosa (McFee et al., 2015). The

visualizations are supported by Matplotlib (Hunter,

2007). Finally, we make use of Numpy for storing

the data during processing (Harris et al., 2020).

If the proposed side channel is viable, we should

expect to see an easily distinguishable pattern of high

intensity, followed by low intensity, at around the 30-

35kHz range, correlating with the idle periods, and

high CPU utilization periods respectively.

4.3 Ideal Scenario

4.3.1 Binary State Identiﬁcation

Our ﬁrst investigation is to test if a side channel ex-

ists under ideal circumstances. For this test case, one

server with one virtual machine attempts to generate

a square wave with a period of two seconds. The shell

script repeats on loop a sequence of a 1 second stress

test, at 100% CPU utilization, followed by a 1 sec-

ond period of inactivity. This is repeated while the

data collection script records the signal in millivolts.

Since the frequency spectrum was affected in the 30-

35 kHz range, we use a sample rate of 100 kHz. The

data is collected over a windows of 30 seconds. After

applying the data processing steps, the spectrogram is

generated as shown in Figure 4.

Figure 4: Spectrogram of square wave with a period of two

seconds.

This spectrogram corroborates our hypothesis that

data can be ex-ﬁltrated. While the machine is in an

Data Leakage in Isolated Virtualized Enterprise Computing Systems

121

idle state, we observe bands of high intensity in the

30-40 kHz range. While the machine is under high

stress, the high intensity region fades to match the

background. An interesting observation is that the

high and low states are recognizable not only between

30 and 40 kHz, but also quite clearly at 10 kHz, al-

though the bands are not as intense. This is impor-

tant to note because it means that a high frequency

sampling circuit is not strictly necessary to be able to

observe the power consumption EMI ﬁngerprint.

For the sake of demonstrating reproducibility, the

experiment is repeated in the same fashion, except

with the square wave having a period of ten seconds

(ﬁve seconds low, ﬁve seconds high). Figure 5 shows

the spectrogram generated from the process.

Figure 5: Spectrogram of square wave with a period of ten

seconds.

The spectrogram shows once again that the square

wave is easily identiﬁable, both between 30 and 40

kHz, and at 10 kHz. Thus we conclude that under

minimal background noise, this attack is discernible.

We now turn our attention to non-binary ex-ﬁltration,

using the CPU usage to inﬂuence the EMI signature.

4.3.2 Ternary State Identiﬁcation

Finally, we tested how different levels of CPU utiliza-

tion would show up in the spectrogram. In stress-ng,

this is accomplished using an argument to specify a

particular CPU utilization level. The shell script is

modiﬁed to cycle from idle, to 50% CPU utilization,

to 100% CPU utilization, repetitively. We expect this

to show three different levels of intensity in the spec-

trogram, which could be used as three separate data

symbols in a streaming information side-channel.

The spectrogram in Figure 6 shows the processed

thirty second trace. The goal in this test was to gener-

ate discernible states of low, medium, and high inten-

sity. As in the prior tests, the low and high states are

quite different visually. However, the medium state is

Figure 6: Spectrogram showing power modulations at three

levels.

not as clearly deﬁned. Interestingly, the pattern does

not manifest as clearly at the 10 kHz range. It is un-

clear if this ternary state can be easily detected auto-

matically. For instance, a machine learning algorithm

may be able to discern the different states from one

another.

5 CONCLUSIONS

Using EMI ﬁngerprints as a side channel for power

analysis in a computing system was shown to be fea-

sible. In particular, we show that binary information

can be ex-ﬁltrated from the system at a rate of at least

1 symbol per second. The receiver of the information

does not need to be co-located with the system un-

der attack—it only requires to be on the same power

line as the computing system (such as a power outlet

in a nearby room). Using a frequency based analy-

sis of the data, a clear signal can be received using

binary symbols. Additional data symbols may also

be possible, but require further investigation. The ex-

perimental results in this paper indicate our position

that network-isolated virtual processes are capable of

leaking sensitive information via a power-based side

channel.

5.1 Background Noise Conditions

While our experiments show promise as a proof of

concept, our test cases represent somewhat ideal cir-

cumstances that are not typical of an enterprise com-

puting environment. In a real-world application, a

server rack might have many servers with multiple

virtual machines running, each performing different

tasks with varying power consumption requirements.

This would generate a signiﬁcant amount of back-

ICISSP 2023 - 9th International Conference on Information Systems Security and Privacy

122

ground noise, which could make the side channel in-

feasible, or at least far more difﬁcult to interpret.

We have performed initial testing with some back-

ground noise conditions, and the results are visually

quite similar to the ideal case. However, a far more

extensive exploration should be done before draw-

ing any conclusions about the effectiveness of this

side channel in a typical enterprise computing envi-

ronment with multiple servers running. To explore

this further, we intend to repeat similar experiments

with various tasks running in the background on vir-

tual machines on other servers. By steadily increasing

the background noise during the experiment, we can

identify at what point the side channel becomes sig-

niﬁcantly noisier, or altogether infeasible.

5.2 Covert Channel

Related to side channel investigations, future work

could also explore the viability of creating a covert

communication channel via power analysis. Given

the experimental results achieved so far, it is possi-

ble that information could be encoded using the high

and low states seen in the spectrogram. With an ap-

propriate encoding scheme, a relatively efﬁcient sys-

tem could be designed to covertly transmit informa-

tion by modulating the power consumption such that

the high and low states represent binary data. The

communication rate could be enhanced by using a

ternary encoding scheme, where the power consump-

tion is modulated between three states, corresponding

to: idle, 50%, and 100% CPU utilization. While we

did perform some testing of this idea (as seen in ﬁgure

6), we intend to design further experiments, and in-

vestigate encoding schemes and data processing tech-

niques for extracting information from the signal, in

order to more thoroughly investigate the viability of

the concept.

REFERENCES

Chen, K.-Y., Cohn, G. A., Gupta, S., and Patel, S. N. (2013).

utouch: sensing touch gestures on unmodiﬁed lcds.

In Proceedings of the SIGCHI Conference on Human

Factors in Computing Systems, pages 2581–2584.

Enev, M., Gupta, S., Kohno, T., and Patel, S. N. (2011).

Televisions, video privacy, and powerline electromag-

netic interference. In Proceedings of the 18th ACM

conference on Computer and communications secu-

rity, pages 537–550.

Golder, A., Das, D., Danial, J., Ghosh, S., Sen, S., and Ray-

chowdhury, A. (2019). Practical approaches toward

deep-learning-based cross-device power side-channel

attack. IEEE Transactions on Very Large Scale Inte-

gration (VLSI) Systems, 27(12):2720–2733.

Gupta, S., Chen, K.-Y., Reynolds, M. S., and Patel, S. N.

(2011). Lightwave: using compact ﬂuorescent lights

as sensors. In Proceedings of the 13th international

conference on Ubiquitous computing, pages 65–74.

Gupta, S., Reynolds, M. S., and Patel, S. N. (2010). Elec-

trisense: Single-point sensing using emi for electri-

cal event detection and classiﬁcation in the home. In

Proceedings of the 12th ACM International Confer-

ence on Ubiquitous Computing, UbiComp ’10, page

139–148, New York, NY, USA. Association for Com-

puting Machinery.

Harris, C. R., Millman, K. J., van der Walt, S. J., Gommers,

R., Virtanen, P., Cournapeau, D., Wieser, E., Taylor,

J., Berg, S., Smith, N. J., Kern, R., Picus, M., Hoyer,

S., van Kerkwijk, M. H., Brett, M., Haldane, A., del

ıo, J. F., Wiebe, M., Peterson, P., G

erard-Marchant,

P., Sheppard, K., Reddy, T., Weckesser, W., Abbasi,

H., Gohlke, C., and Oliphant, T. E. (2020). Array pro-

gramming with NumPy. Nature, 585(7825):357–362.

Hunter, J. D. (2007). Matplotlib: A 2d graphics environ-

ment. Computing in Science & Engineering, 9(3):90–

95.

Khan, H. A., Sehatbakhsh, N., Nguyen, L. N., Prvulovic,

M., and Zaji

c, A. (2019). Malware detection in em-

bedded systems using neural network model for elec-

tromagnetic side-channel signals. Journal of Hard-

ware and Systems Security, 3(4):305–318.

Kocher, P., Jaffe, J., and Jun, B. (1999). Differential power

analysis. In Wiener, M., editor, Advances in Cryptol-

ogy — CRYPTO’ 99, pages 388–397, Berlin, Heidel-

berg. Springer Berlin Heidelberg.

Lipp, M., Kogler, A., Oswald, D., Schwarz, M., Easdon,

C., Canella, C., and Gruss, D. (2021). Platypus:

Software-based power side-channel attacks on x86. In

2021 IEEE Symposium on Security and Privacy (SP),

pages 355–371.

McFee, B., Raffel, C., Liang, D., Ellis, D. P. W., McVicar,

M., Battenberg, E., and Nieto, O. (2015). librosa: Au-

dio and music signal analysis in python.

Patel, S. N., Robertson, T., Kientz, J. A., Reynolds, M. S.,

and Abowd, G. D. (2007). At the ﬂick of a switch: De-

tecting and classifying unique electrical events on the

residential power line (nominated for the best paper

award). In International Conference on Ubiquitous

Computing, pages 271–288. Springer.

Szefer, J. (2019). Survey of microarchitectural side and

covert channels, attacks, and defenses. Journal of

Hardware and Systems Security, 3(3):219–234.

Taylor, M. A., Larson, E. C., and Thornton, M. A. (2021).

Rapid ransomware detection through side channel ex-

ploitation. In 2021 IEEE International Conference on

Cyber Security and Resilience (CSR), pages 47–54.

Data Leakage in Isolated Virtualized Enterprise Computing Systems

123