Remote physical device fingerprinting

Tadayoshi Kohno

CSE Department, UC San Diego

tkohno@cs.ucsd.edu

Andre Broido

CAIDA, UC San Diego

broido@caida.org

kc claffy

CAIDA, UC San Diego

kc@caida.org

Abstract

We introduce the area of remote physical device finger-

printing, or fingerprinting a physical device, as opposed to

an operating system or class of devices, remotely, and with-

out the fingerprinted device’s known cooperation. We ac-

complish this goal by exploiting small, microscopic devia-

tions in device hardware: clock skews. Our techniques do

not require any modification to the fingerprinted devices.

Our techniques report consistent measurements when the

measurer is thousands of miles, multiple hops, and tens of

milliseconds away from the fingerprinted device, and when

the fingerprinted device is connected to the Internet from

different locations and via different access technologies.

Further, one can apply our passive and semi-passive tech-

niques when the fingerprinted device is behind a NAT or

firewall, and also when the device’s system time is main-

tained via NTP or SNTP. One can use our techniques to

obtain information about whether two devices on the Inter-

net, possibly shifted in time or IP addresses, are actually the

same physical device. Example applications include: com-

puter forensics; tracking, with some probability, a physical

device as it connects to the Internet from different public ac-

cess points; counting the number of devices behind a NAT

even when the devices use constant or random IP IDs; re-

motely probing a block of addresses to determine if the ad-

dresses correspond to virtual hosts, e.g., as part of a virtual

honeynet; and unanonymizing anonymized network traces.

1 Introduction

There are now a number of powerful techniques for re-

mote operating system fingerprinting, i.e., techniques for

remotely determining the operating systems of devices on

the Internet [2, 3, 5, 27]. We push this idea further and in-

troduce the notion of remote physical device fingerprinting,

or remotely fingerprinting a physical device, as opposed to

an operating system or class of devices, without the finger-

printed device’s known cooperation. We accomplish this

goal to varying degrees of precision by exploiting micro-

scopic deviations in device hardware: clock skews.

CLASSES OF FINGERPRINTING TECHNIQUES. We con-

sider three main classes of remote physical device finger-

printing techniques: passive, active, and semi-passive. The

first two have standard definitions — to apply apassive

fingerprinting technique, the fingerprinter (measurer,at-

tacker,adversary) must be able to observe traffic from the

device (the fingerprintee) that the attacker wishes to finger-

print, whereas to apply an active fingerprinting technique,

the fingerprinter must have the ability to initiate connec-

tions to the fingerprintee. Our third class of techniques,

which we refer to as semi-passive fingerprinting techniques,

assumes that after the fingerprintee initiates a connection,

the fingerprinter has the ability to interact with the finger-

printee over that connection; e.g., the fingerprinter is a web-

site with which the device is communicating, or is an ISP

in the middle capable of modifying packets en route. Each

class of techniques has its own advantages and disadvan-

tages. For example, passive techniques will be completely

undetectable to the fingerprinted device, passive and semi-

passive techniques can be applied even if the fingerprinted

device is behind a NAT or firewall, and semi-passive and

active techniques can potentially be applied over longer pe-

riods of time; e.g., after a laptop connects to a website and

the connection terminates, the website can still continue to

run active measurements.

METHODOLOGY. For all our methods, we stress that the

fingerprinter does not require any modification to or co-

operation from the fingerprintee; e.g., we tested our tech-

niques with default Red Hat 9.0, Debian 3.0, FreeBSD

5.2.1, OpenBSD 3.5, OS X 10.3.5 Panther, Windows XP

SP2, and Windows for Pocket PC 2002 installations.1In Ta-

ble 1 we summarize our preferred methods for fingerprint-

ing the most popular operating systems.

Our preferred passive and semi-passive techniques ex-

ploit the fact that most modern TCP stacks implement the

1Our techniques work for the default installs of other versions of these

operating systems; here we just mention the most recent stable versions of

the operating systems that we analyzed.

Technique and section Class NTP Red Hat 9.0 OS X Panther Windows XP

TCP timestamps, Section 3 passive Yes Yes Ye s No

TCP timestamps, Section 3 semi-passive Ye s Ye s Ye s Ye s

ICMP tstamp requests, Section 4 active No Yes No Ye s

Table 1. This table summarizes our main clock skew-based physical device fingerprinting techniques.

A “Yes” in the NTP column means that one can use the attack regardless of whether the fingerprintee

maintains its system time with NTP [19]. One can use passive and semi-passive techniques when

the fingerprintee is behind a NAT or current generation firewall.

TCP timestamps option from RFC 1323 [13] whereby, for

performance purposes, each party in a TCP flow includes

information about its perception of time in each outgoing

packet. A fingerprinter can use the information contained

within the TCP headers to estimate a device’s clock skew

and thereby fingerprint a physical device. We stress that

one can use our TCP timestamps-based method even when

the fingerprintee’s system time is maintained via NTP [19].

While most modern operating systems enable the TCP

timestamps option by default, Windows 2000 and XP ma-

chines do not. Therefore, we developed a trick, which in-

volves an intentional violation of RFC 1323 on the part of

a semi-passive or active adversary, to convince Microsoft

Windows 2000 and XP machines to use the TCP times-

tamps option in Windows-initiated flows. In addition to

our TCP timestamps-based approach, we consider passive

fingerprinting techniques that exploit the difference in time

between how often other periodic activities are supposed to

occur and how often they actually occur, and we show how

one might use a Fourier transform on packet arrival times

to infer a device’s clock skew. Since we believe that our

TCP timestamps-based approach is currently our most gen-

eral passive technique, we focus on the TCP timestamps

approach in this paper.

An active adversary could also exploit the ICMP proto-

col to fingerprint a physical device. Namely, an active ad-

versary could issue ICMP Timestamp Request messages to

the fingerprintee and record a trace of the resulting ICMP

Timestamp Reply messages. If the fingerprintee does not

maintain its system time via NTP or does so only infre-

quently and if the fingerprintee replies to ICMP Timestamp

Requests, then an adversary analyzing the resulting ICMP

Timestamp Reply messages will be able to estimate the fin-

gerprintee’s system time clock skew. Default Red Hat 9.0,

Debian 3.0, FreeBSD 5.2.1, OpenBSD 3.5, and Windows

2000 and XP and Pocket PC 2002 installations all satisfy

the above preconditions.

PARAMETERS OF INVESTIGATION. Toward developing the

area of remote physical device fingerprinting via remote

clock skew estimation, we must address the following set

of interrelated questions:

(1) For what operating systems are our remote clock skew

estimation techniques applicable?

(2) What is the distribution of clock skews across multiple

fingerprintees? And what is the resolution of our clock

skew estimation techniques? (I.e., can one expect two

machines to have measurably different clock skews?)

(3) For a single fingerprintee, can one expect the clock

skew estimate of that fingerprintee to be relatively

constant over long periods of time, and through re-

boots, power cycles, and periods of down time?

(4) What are the effects of a fingerprintee’s access tech-

nology (e.g., wireless, wired, dialup, cable modem)

on the clock skew estimates for the device?

(5) How are the clock skew estimates affected by the dis-

tance between the fingerprinter and the fingerprintee?

(6) Are the clock skew estimates independent of the fin-

gerprinter? I.e., when multiple fingerprinters are mea-

suring a single fingerprintee at the same time, will they

all output (approximately) the same skew estimates?

(7) How much data do we need to be able to remotely

make accurate clock skew estimates?

Question (6) is applicable because common fingerprinters

will probably use NTP-based time synchronization when

capturing packets, as opposed to more precise CDMA- or

GPS-synchronized timestamps. Answers to the above ques-

tions will help determine the efficacy of our physical device

fingerprinting techniques.

EXPERIMENTS AND HIGH-LEVEL RESULTS. To under-

stand and refine our techniques, we conducted experiments

with machines that we controlled and that ran a variety of

operating systems, including popular Linux, BSD, and Mi-

crosoft distributions. In all cases we found that we could

use at least one of our techniques to estimate clock skews

of the machines, and that we required only a small amount

of data, though the exact data requirements depended on the

operating system in question. For the most popular operat-

ing systems, we observed that when the system did not use

NTP- or SNTP-based time synchronization, then the TCP

timestamps-based and the ICMP-based techniques yielded

approximately the same skew estimates. This result, cou-

pled with details that we describe in the body, motivated

us to use the TCP timestamps-based method in most of our

experiments. We survey some of our experiments here.

To understand the effects of topology and access tech-

nology on our skew estimates, we fixed the location of the

fingerprinter and applied our TCP timestamps-based tech-

nique to a single laptop in multiple locations, on both North

American coasts, from wired, wireless, and dialup loca-

tions, and from home, business, and campus environments

(Table 3). All clock skew estimates for the laptop were

close — the difference between the maximum and the min-

imum skew estimate was only 0.67 ppm. We also simul-

taneously measured the clock skew of the laptop and an-

other machine from multiple PlanetLab nodes throughout

the world, as well as from a machine of our own with a

CDMA-synchronized Dag card [1, 9, 11, 17] for taking net-

work traces with precise timestamps (Table 4). With the ex-

ception of the measurements taken by a PlanetLab machine

in India (over 300 ms round trip time away), for each exper-

iment, all the fingerprinters (in North America, Europe, and

Asia) reported skew estimates within only 0.56 ppm of each

other. These experiments suggest that, except for extreme

cases, the results of our clock skew estimation techniques

are independent of access technology and topology.

Toward understanding the distribution of clock skews

across machines, we applied the TCP timestamps technique

to the devices in a trace collected on one of the U.S.’s Tier 1

OC-48 links (Figure 2). We also measured the clock skews

of 69 (seemingly) identical Windows XP SP1 machines in

one of our institution’s undergraduate computing facilities

(Figure 3). The latter experiment, which ran for 38 days,

as well as other experiments, show that the clock skew es-

timates for any given machine are approximately constant

over time, but that different machines have detectably dif-

ferent clock skews. Lastly, we use the results of these and

other experiments to argue that the amount of data (packets

and duration of data) necessary to perform our skew estima-

tion techniques is low, though we do not perform a rigorous

analysis of exactly what “low” means.

APPLICATIONS AND ADDITIONAL EXPERIMENTS.Totest

the applicability of our techniques, we applied our tech-

niques to a honeyd [24] virtual honeynet consisting of 100

virtual Linux 2.4.18 hosts and 100 virtual Windows XP SP1

hosts. Our experiments showed with overwhelming proba-

bility that the TCP flows and ICMP timestamp responses

were all handled by a single machine as opposed to 200

different machines. We also applied our techniques to a

network of five virtual machines running under VMware

Workstation [4] on a single machine. In this case, the clock

skew estimates of the virtual machines are significantly dif-

ferent from what one would expect from real machines (the

skews were large and not constant over time; Figure 5). An

application of our techniques, or natural extensions, might

therefore be to remotely detect virtual honeynets.

Another applications of our techniques is to count the

number of hosts behind a NAT, even if those hosts use ran-

dom or constant IP IDs to counter Bellovin’s attack [7],

even if all the hosts run the same operating system, and even

if not all of the hosts are up at the same time. Furthermore,

when both our techniques and Bellovin’s techniques are ap-

plicable, we expect our approach to provide a much higher

degree of resolution. One could also use our techniques for

forensics purposes, e.g., to argue whether or not a given lap-

top was connected to the Internet from a given access loca-

tion. One could also use our techniques to help track laptops

as they move, perhaps as part of a Carnivore-like project

(here we envision our skew estimates as one important com-

ponent of the tracking; other components could be informa-

tion gleaned from existing operating system fingerprinting

techniques, usage characteristics, and other heuristics). One

can also use our techniques to catalyze the unanonymization

of prefix-preserving anonymized network traces [28, 29].

BACKGROUND AND RELATED WORK. It has long been

known that seemingly identical computers can have dis-

parate clock skews. The NTP [19] specification describes

a method for reducing the clock skews of devices’ sys-

tem clocks, though over short periods of time an NTP-

synchronized machine may still have slight clock skew. In

1998 Paxson [22] initiated a line of research geared toward

eliminating clock skew from network measurements, and

one of the algorithms we use is based on a descendent of

the Paxson paper by Moon, Skelly, and Towsley [20]. Fur-

ther afield, though still related to clock skews, P´asztor and

Veitch [21] have created a software clock on a commod-

ity PC with high accuracy and small clock skew. One fun-

damental difference between our work and previous work

is our goal: whereas all previous works focus on creat-

ing methods for eliminating the effects of clock skews, our

work exploits and capitalizes on the effects of clock skews.

Anagnostakis et. al. [6] use ICMP Timestamp Requests

to study router queuing delays. It is well known that a net-

work card’s MAC address is supposed to be unique and

therefore could serve as a fingerprint of a device assum-

ing that the adversary can observe the device’s MAC ad-

dress and that the owner of the card has not changed the

MAC address. The main advantage of our techniques over

a MAC address-based approach is that our techniques are

mountable by adversaries thousands of miles and multiple

hops away. One could also use cookies or any other per-

sistent data to track a physical device, but such persistent

data may not always be available to an adversary, perhaps

because the user is privacy-conscious and tries to minimize

storage and transmission of such data, or because the user

never communicates that data unencrypted.

See [15] for the full version of this paper.

2 Clocks and clock skews

When discussing clocks and clock skews, we build on

the nomenclature from the NTP specification [19] and from

Paxson [22]. A clock Cis designed to represent the amount

of time that has passed since some initial time i[C].Clock

C’s resolution,r[C], is the smallest unit by which the clock

can be incremented, and we refer to each such increment

as a tick. A resolution of 10 ms means that the clock is de-

signed to have 10 ms granularity, not that the clock is always

incremented exactly every10ms. ClockC’s intended fre-

quency,Hz[C], is the inverse of its resolution; e.g., a clock

with 10 ms granularity is designed to run at 100 Hz. For

all t≥i[C],letR[C](t)denote the time reported by clock

Cat time t,wheretdenotes the true time as defined by

national standards. The offset of clock C,oﬀ[C], is the dif-

ference between the time reported by Cand the true time,

i.e., oﬀ[C](t)=R[C](t)−tfor all t≥i[C]. A clock’s

skew,s[C], is the first derivative of its offset with respect to

time, where we assume for simplicity of notation that R[C]

is a differentiable function in t. We report skew estimates in

microseconds per second (µs/s) or, equivalently, parts per

million (ppm). As we shall show, and as others have also

concluded [22, 20, 26], it is often reasonable to assume that

a clock’s skew is constant. When the clock in question is

clear from context, we shall remove the parameter Cfrom

our notation; e.g., s[C]becomes s.

A given device can have multiple, possibly independent,

clocks. For remote physical device fingerprinting, we ex-

ploit two different clocks: the clock corresponding to a de-

vice’s system time, and a clock internal to a device’s TCP

network stack, which we call the device’s TCP timestamps

option clock or TSopt clock. We do not consider the hard-

ware bases for these clocks here since our focus is not on

understanding why these clocks have skews, but on exploit-

ing the fact these clocks can have measurable skews on pop-

ular current-generation systems.

THE SYSTEM CLOCK. To most users of a computer sys-

tem, the most visible clock is the device’s system clock,

Csys, which is designed to record the amount of time since

00:00:00 UTC, January 1, 1970. Although the system

clocks on professionally administered machines are often

approximately synchronized with true time via NTP [19]

or, less accurately, via SNTP [18], we stress that it is much

less likely for the system clocks on non-professionally man-

aged machines to be externally synchronized. This lack of

synchronization is because the default installations of most

of the popular operating systems that we tested do not syn-

chronize the hosts’ system clocks with true time or, if they

do, they do so only infrequently. For example, default Win-

dows XP Professional installations only synchronize their

system times with Microsoft’s NTP server when they boot

and once a week thereafter. Default Red Hat 9.0 Linux

installations do not use NTP by default, though they do

present the user with the option of entering an NTP server.

Default Debian 3.0, FreeBSD 5.2.1, and OpenBSD 3.5 sys-

tems, at least under the configurations that we selected (e.g.,

“typical user”), do not even present the user with the op-

tion of installing ntpd. For such a non-professionally-

administered machine, if an adversary can learn the values

of the machine’s system clock at multiple points in time,

the adversary will be able to infer information about the de-

vice’s system clock skew,s[Csys].

THE TCP TIMESTAMPS OPTION CLOCK. RFC 1323 [13]

specifies the TCP timestamps option to the TCP protocol.

A TCP flow will use the TCP timestamps option if the net-

work stacks on both ends of the flow implement the option

and if the initiator of the flow includes the option in the

initial SYN packet. All modern operating systems that we

tested implement the TCP timestamps option. Of the sys-

tems we tested, Microsoft Windows 2000 and XP are the

only ones that do not include the TCP timestamps option in

the initial SYN packet (Microsoft Windows Pocket PC 2002

does include the option when initiating TCP flows). In Sec-

tion 3 we introduce a trick for making Windows 2000- and

XP-initiated flows use the TCP timestamps option.

For physical device fingerprinting, the most important

property of the TCP timestamps option is that if a flow uses

the option, then a portion of the header of each TCP packet

in that flow will contain a 32-bit timestamp generated by

the creator of that packet. The RFC does not dictate what

values the timestamps should take, but does say that the

timestamps should be taken from a “virtual clock” that is “at

least approximately proportionalto real time [13];” the RFC

1323 PAWS algorithm does stipulate (Section 4.2.2) that the

resolution of this virtual clock be between 1 ms and 1 sec-

ond. We refer to this “virtual clock” as the device’s TCP

timestamps option clock, or its TSopt clock Ctcp . There is no

requirement that a device’s TSopt clock and its system clock

be correlated. Moreover, for popular operating systems like

Windows XP, Linux, and FreeBSD, a device’s TSopt clock

may be unaffected by adjustments to the device’s system

clock via NTP. To sample some popular operating systems,

standard Red Hat 9.0 and Debian 3.0 Linux distributions2

and FreeBSD 5.2.1 machines have TSopt clocks with 10 ms

resolution, OS X Panther and OpenBSD 3.5 machines have

TSopt clocks with 500 ms resolution, and Microsoft Win-

dows 2000, XP, and Pocket PC 2002 systems have TSopt

clocks with 100 ms resolution. Most systems reset their

TSopt clock to zero upon reboot; on these systems i[Ctcp]

is the time at which the system booted. If an adversary can

learn the values of a device’s TSopt clock at multiple points

in time, then the adversary may be able to infer information

about the device’s TSopt clock skew,s[Ctcp].

2We do not generalize this to all Linux distributions since Knoppix 3.6,

with the 2.6.7 experimental kernel, has 1 ms resolution.

3 Exploiting the TCP Timestamps Option

In this section we consider (1) how an adversary might

obtain samples of a device’s TSopt clock at multiple points

in time and (2) how an adversary could use those samples

to fingerprint a physical device. We assume for now that

there is a one-to-one correspondence between physical de-

vices and IP addresses, and defer to Section 8 a discussion

of how to deal with multiple active hosts behind a NAT; in

this section we do consider NATs with a single active device

behind them.

THE MEASURER. The measurer can be any entity capable

of observing TCP packets from the fingerprintee, assum-

ing that those packets have the TCP timestamps option en-

abled. The measurer could therefore be the fingerprintee’s

ISP, or any tap in the middle of the network over which

packets from the device travel; e.g., we apply our techniques

to a trace taken on a major Tier 1 ISP’s backbone OC-48

links. The measurer could also be any system with which

the fingerprintee frequently communicates; prime examples

of such systems include a search engine like Google, a news

website, and a click-through ads service that displays con-

tent on a large number of websites. If the measurer is ac-

tive, then the measurer could also be the one to initiate a

TCP flow with the fingerprintee, assuming that the device

is reachable and has an open port. If the measurer is semi-

passive or active, then it could make the flows that it ob-

serves last abnormally long, thereby giving the measurer

samples of the fingerprintee’s clock over extended periods

of time.

ATRICK FOR MEASURING WINDOWS 2000 AND XP MA-

CHINES. We seek the ability to measure TSopt clock skews

of Windows 2000 and XP machines even if those machines

are behind NATs and firewalls. But, because of the nature of

NATs and firewalls, in these cases we will typically be lim-

ited to analyzing flows initiated by the Windows machines.

Unfortunately, because Windows 2000 and XP machines do

not include the TCP timestamps option in their initial SYN

packets, the TCP timestamps RFC [13] mandates that none

of the subsequent packets in Windows-initiated flows can

include the TCP timestamps option. Thus, assuming that all

parties correctly implement the TCP RFCs, a passive adver-

sary will not be able to exploit the TCP timestamps option

with Windows 2000/XP-initiated flows.

If the adversary is semi-passive, we observe the follow-

ing trick. Assume for simplicity that the adversary is the de-

vice to whom the Windows machine is connecting. After re-

ceiving the initial SYN packet from the Windows machine,

the adversary will reply with a SYN/ACK, but the adversary

will break the RFC 1323 specification and include the TCP

timestamps option in its reply. After receiving such a reply,

our Windows 2000 and XP machines ignored the fact that

they did not include the TCP timestamps option in their ini-

tial SYN packets, and included the TCP timestamps option

in all of their subsequent packets. As an extension, we note

that the adversary does not have to be the device to whom

the Windows machine is connecting. Rather, the adversary

simply needs to be able to mount a “device-in-the-middle”

attack and modify packets such that the Windows machine

receives one with the TCP timestamps option turned on. If

the adversary is the device’s ISP, then the ISP could rewrite

the Windows machine’s initial SYN packets so that they in-

clude the TCP timestamps option. The SYN/ACKs from

the legitimate recipients will therefore have the TCP times-

tamps option enabled and, from that point forward, the Win-

dows machine will include the TCP timestamps option in all

subsequent packets in the flows.

We applied this technique to Windows XP machines on

a residential cable system with a LinkSys Wireless Access

Point and a NAT, as well as to Windows XP SP2 machines

using the default XP SP2 firewall, and to Windows XP SP1

machines with the Windows ZoneAlarm firewall. (While

current firewalls do not detect this trick, it is quite possible

that future firewalls might.)

ESTIMATING THE TSOPT CLOCK SKEW. Let us now as-

sume that an adversary has obtained a trace Tof TCP pack-

ets from the fingerprintee, and let us assume for simplicity

that all |T | packets in the trace have the TCP timestamps

option enabled. Toward estimating a device’s TSopt clock

skew s[Ctcp]we adopt the following additional notation. Let

tibe the time in seconds at which the measurer observed the

i-th packet in Tand let Tibe the Ctcp timestamp contained

within the i-th packet. Define

xi=ti−t1

vi=Ti−T1

wi=vi/Hz

yi=wi−xi

OT={(xi,y

i):i∈{1,...,|T |} } .

The unit for wiis seconds; yiis the observed offset of the i-

th packet; OTis the the offset-set corresponding to the trace

T. We discuss below how to compute Hz if it is not known

to the measurer in advance. As an example, Figure 1 shows

the offset-sets for two devices in a two-hour trace of traffic

from an Internet backbone OC-48 link on 2004-04-28 (we

omit IP addresses for privacy reasons). Shifting the clocks

by t1and T1for xiand viis not necessary for our analysis

but makes plots like in Figure 1 cleaner.

If we could assume that the measurer’s clock is accurate

and that the tvalues represent true time, and if we could as-

sume that there is no delay between when the fingerprintee

generates the i-th packet and when the measurer records

the i-th packet, then yi=oﬀ(xi+t1). Under these as-

sumptions, and if we make the additional assumption that

0 900 1800 2700 3600 4500 5400 6300 7200

time since start of measurement (seconds)

-600

-400

-200

200

400

observed offset (ms)

Source 1: 10Hz TSopt clock, 37526 packets, ttl=113

Source 2: 100Hz TSopt clock, 20974 packets, ttl=55

linear programming-based upper bound

Figure 1. TSopt clock offset-sets for two

sources in BBN. Trace recorded on an OC-

48 link of a U.S. Tier 1 ISP, 2004-04-28 19:30–

21:30PDT. The source with the wide band has

a 10 Hz TSopt clock, the source with the nar-

row band has a 100 Hz TSopt clock. A source

with no clock skew would have a horizontal

band.

Ris differentiable, then the first derivative of y,whichis

the slope of the points in OT,istheskewsof Ctcp .Since

we cannot generally make these assumptions, we are left to

approximate sfrom the data.

Let us consider plots like those in Figure 1 more closely.

We first observe that the large band corresponds to a device

where the TSopt clock has low resolution (r= 100 ms) and

that the narrow band corresponds to a device with a higher

resolution (r=10ms). The width of these bands, and in

particular the wide band, means that if the duration of our

trace is short, we cannot always approximate the slope of

the points in OTby computing the slope between any two

points in the set. Moreover, as Paxson and others have noted

in similar contexts [22, 20], variable network delay renders

simple linear regression insufficient. Consequently, to ap-

proximate the the skew sfrom OT, we borrow a linear pro-

gramming solution from Moon, Skelly, and Towsley [20],

which has as its core Graham’s convex hull algorithm on

sorted data [12].

The linear programming solution outputs the equation of

a line αx +βthat upper-bounds the set of points OT.We

use an upper bound because network and host delays are all

positive. The slope of the line, α, is our estimate of the clock

skew of Ctcp. In detail, the linear programming constraints

for this line are that, for all i∈{1,...,|T |},

α·xi+β≥yi,

which means that the solution must upper-bound all the

points in OT. The linear programming solution then mini-

mizes the average vertical distance of all the points in OT

from the line; i.e., the linear programming solution is one

that minimizes the objective function

|T | ·

|T |



i=1

α·xi+β−yi.

Although one can solve the above using standard linear pro-

gramming techniques, as Moon, Skelly, and Towsley [20]

note, there exist techniques to solve linear programming

problems in two variables in linear time [10, 16]. We use

a linear time algorithm in all our computations.

It remains to discuss how to infer Hz if the measurer does

not know it in advance. One solution involves computing

the slope of the points

I={(xi,v

i):i∈{1,...,|T | }

and rounding to the nearest integer. One can compute the

slope of this set by adapting the above linear programming

problem to this set.

AN EQUIVALENT VIEW.IfAis the slope of the points in

the above set I, derived using the linear programming al-

gorithm, then one could also approximate the skew of Ctcp

as A/Hz −1. This approach is simply a different way of

arriving at the same solution since we can prove that, when

using the linear programming method for slope estimation,

both approaches produce the same skew estimate. We use

the offset-set approach since these sets naturally yield fig-

ures where the skews are clearly visible; e.g., Figure 1.

4 Exploiting ICMP Timestamp Requests

THE MEASURER. To exploit a device’s system time clock

skew, the measurer could be any website with which the fin-

gerprintee communicates, or any other device on the Inter-

net provided that the measurer is capable of issuing ICMP

Timestamp Requests (ICMP message type 13) to the fin-

gerprintee. The measurer must also be capable of record-

ing the fingerprintee’s subsequent ICMP Timestamp Reply

messages (ICMP message type 14). In order for this tech-

nique to be mountable, the primary limitation is that the de-

vice must not be behind a NAT or firewall that filters ICMP.

ESTIMATING THE SYSTEM CLOCK SKEW. Let us now as-

sume that an adversary has obtained a trace Tof ICMP

Timestamp Reply messages from the fingerprintee. The

ICMP Timestamp Reply messages will contain two 32-bit

values generated by the fingerprintee. The first value is

the time at which the corresponding ICMP Timestamp Re-

quest packet was received, and the second value is the time

at which the ICMP Timestamp Reply was generated; here

time is according to the fingerprintee’s system clock, Csys ,

and is reported in milliseconds since midnight UTC. Win-

dows machines report the timestamp in little endian for-

mat, whereas all the other machines that we tested report

the timestamp in big endian notation. The remaining nota-

tion and the method for skew estimation is now identical to

what we presented in Section 3, with two minor exceptions.

First, the adversary does not have to compute Hz since RFC

792 [23] requires that Hz be 1000 (or, if not, that a spe-

cial bit be set to indicate non-compliance). Second, since

the time reported in the ICMP Timestamp Reply is in mil-

liseconds since midnight UTC, we expect the timestamps

reported in the ICMP Timestamp Reply messages to reset

approximately once a day; we adjust the vvalues accord-

ingly. (The only thing special that our attack exploits about

ICMP is the fact that ICMP has a message type that will

reveal a device’s system time; our techniques would work

equally well with any other protocol that leaks information

about a device’s system or other clock.)

BRIEF COMPARISON WITH TCP TIMESTAMPS.Formuch

of the rest of this paper, we focus on our TCP timestamps-

based approach for physical device fingerprinting rather

than our ICMP-based approach. This is not because we

consider the ICMP-based approach to be inferior. Rather,

we focus on the TCP timestamps-based approach because

most systems have TSopt clocks that operate at a lower

frequency than the 1000 Hz clocks included in the ICMP

timestamp reply messages. This means that it should re-

quire less data for an active adversary to mount our ICMP

fingerprinting technique than to mount our TCP timestamps

technique. Our positive results for the TCP timestamps-

based fingerprinting techniques imply that the ICMP-based

fingerprinting technique will be effective and will have low

data requirements. Focusing on our TCP timestamps based

approach also allows us to experiment with machines be-

hind NATs and firewalls. We also remark that for popular

operating systems, if a system does not externally synchro-

nize its system time, then the system’s TSopt and system

clocks will be highly correlated (Section 7), which means

that the distribution of system clock skews for machines not

using NTP will be similar to the distribution of TSopt clocks

skews.

5 Distribution and stability of TSopt clock

skew measurements

We now address two fundamental properties that must

hold in order for remote clock skew estimation to be an

effective physical device fingerprinting technique. First,

we show that there is variability in different devices’ clock

skews, meaning that it is reasonable to expect different de-

vices on the Internet to have measurably different clock

skews. Second, we give evidence to suggest that clock

skews, as measured by our techniques, are relatively con-

stant over time. These two facts provide the basis for our

min pkts min duration total sources with entropy

per hour per hour sources stable skews (bits)

(mp)(md,mins) (|S|) (|S |)

010 18335 8225 4.87

030 13517 6859 5.39

050 7246 4120 5.87

500 10 4356 2583 5.99

500 30 4016 2446 6.11

500 50 3368 2104 6.18

2000 10 1730 1116 6.22

2000 30 1629 1077 6.32

2000 50 1489 1009 6.41

Table 2. Entropy estimates from BB-2004-04-

28 when pv =1ppm.

use of remote clock skew estimation as a physical device

fingerprinting technique since they imply that an adversary

can gain (sometimes significant) information by applying

our techniques to measure a device’s or set of devices’ clock

skews.

The novelty here is not in claiming that these proper-

ties are true. Indeed, it is well known that different com-

puter systems can have different clock skews, and others

[22, 20, 21, 26] have argued that a given device generally

has a constant clock skew. Rather, the contribution here is

showing that these properties survive our remote clock skew

estimation techniques and, in the case of our analyses of the

distribution of clock skews, measuring the bits of informa-

tion (entropy) a passive adversary might learn by passively

measuring the TSopt clock skews of fingerprintees.

DISTRIBUTION OF CLOCK SKEWS:ANALYSISOFPAS-

SIVE TRACES. Our first experiment in this section focuses

on understanding the distribution of clock skews across de-

vices as reported by our TCP timestamps-based passive fin-

gerprinting technique. For this experiment we analyzed a

passive trace of traffic in both directions of a major OC-48

link; CAIDA collected the trace between 19:30 and 21:30

PDT on 2004-04-28. Since the OC-48 link runs North-

South, let BBNdenote the Northbound trace, and let BBS

denote the Southbound trace (BB stands for backbone).

CAIDA obtained the traces using differentDag [11] cards in

each direction; these cards’ clocks were synchronized with

each other, but not with true time. This latter property does

not affect the following discussion because (1) the clock

skews of the Dag cards appear to be constant and therefore

only shift our skew estimates by a constant amount and (2)

here we are only interested in the general distribution of the

clock skews of the sources in the traces.

Let mp and md be positive integers. For simplicity, fix

BB =BBNor BBS. Also assume for simplicity that BB

only contains TCP packets with the TCP timestamps option

turned on. Recall that the trace BB last for two hours. At

a high-level, our analysis considers the set Sof sources in

BB that have ≥mp packets in both the first and the sec-

ond hours, and where the differences in time between the

source’s first and last packets in each hour are ≥md min-

utes. If mp and md are large, then the sources in Sall gen-

erate a large number of packets, and over a long period of

time.

For each source in S, we apply our clock skew estima-

tion technique from Section 3 to the full trace, the first hour

only, and the second hour only. Let pv be a positive number,

and let Sbe the subset of Scorresponding to the sources

whose skew estimates for the full trace, the first hour, and

the second hour are all within pv ppm of each other, and

whose intended frequency Hz is one of the standard val-

ues (1, 2, 10, 100, 512, 1000). If pv is small, then we are

inclined to believe that the skew estimates for the sources

in Sclosely approximate the true skews of the respective

sources. Table 2 shows values of |S| and |S|for different

values of mp and md and when pv =1ppm.

The value |S|/|S | gives an indication of the ratio of

sources of which we can accurately (within pv ppm) mea-

sure the clock skew. While useful, this value provides little

information about the actual distribution of the clock skew

estimates. Much more (visually) telling are images such as

Figure 2, which shows a histogram of the skew estimates

(for the full two hour trace) for all the sources in Swhen

mp = 2000,md =50minutes, and pv =1ppm. (The true

histogram may be shifted horizontally based on the clock

skew of the Dag cards, but a horizontal shift does not af-

fect the general shape of the distribution.) Empirically, for

any given values for mp,md,andpv, we can compute the

entropy of the distribution of clock skews. Doing so serves

as a means of gauging how many bits of information an

adversary might learn by passively monitoring a device’s

clock skew, assuming that devices’ clock skews are con-

stant over time, which is something we address later. To

compute the entropy, we consider bins of width pv,andfor

each source sin S, we increment the count of the bin cor-

responding to devices with clock skews similar to the skew

of s(here we use the skew estimate computed over full two

hours). We then allocate another bin of size |S|−|S|;this

bin counts the number of sources that do not have consis-

tent clock skew measurements. We apply the standard en-

tropy formula [25] to compute the entropy of this distribu-

tion of bins, the results of which appear in the last column

of Table 2. As one might expect, the amount of information

available to an adversary increases as mp and md increase.

Assuming that clock skews are constant over time, our

results suggest that a passive adversary could learn at least

six bits of information about a physical device by applying

our techniques from Section 3. More bits of information

-300 -200 -100 0 100 200 300

skew estimate (ppm)

100

count

Figure 2. Histogram of TSopt clock skew es-

timates for sources in BBN. Trace recorded

on an OC-48 link of a U.S. Tier 1 ISP, 2004-04-

28 19:30–21:30PDT. Here mp = 2000 packets,

md =50minutes, and pv =1ppm.

should be available to an active adversary since an active

adversary might be able to force the fingerprintee to send

packets more frequently or over longer periods of time.

DISTRIBUTION OF CLOCK SKEWS:EXPERIMENTS WITH

A HOMOGENEOUS LAB. One observation on the above

analysis is that we applied it to a wide variety of machines,

which likely ran a wide variety of operating systems. There-

fore, one may wonder whether the distribution shown in

Figure 2 is due to operating system differences or to ac-

tual physical differences on the devices. For example, given

only the above results, it might still be possible to argue

that if we applied our skew estimates to a large number

of (apparently) homogeneous machines, we would get back

approximately the same (i.e., indistinguishable) skew esti-

mates for all of the machines. To address this issue, we con-

ducted an experiment with 69 (apparently) homogeneous

machines in one of UCSD’s undergraduate computing lab-

oratories. All the machines were Micron PCs with 448MHz

Pentium II processors running Microsoft Windows XP Pro-

fessional Service Pack 1. Our measurer, host2,wasa

Dell Precision 410 with a 448MHz Pentium III processor

and running Debian 3.0 with a recompiled 2.4.18 kernel;

host2 is located within the University’s computer science

department and is 3 hops and a half a millisecond away from

the machines in the undergraduate laboratory.

To create the requisite trace of TCP packets from these

machines, we repeatedly opened and then closed connec-

tions from host2 to each of these machines. Each open-

then-close resulted in the Windows machines sending two

packets to host2 with the TCP timestamps option turned

on (the Windows machine sent three packets for each flow,

but the TCP timestamp was always zero in the first of these

three packets). Because of our agreement with the admin-

istrators of these machines, we were only able to open and

close connections with these Windows machines at random

intervals between zero and five minutes long. Thus, on av-

erage we would expect to see each machine send host2

48 TCP packets with the TCP timestamps option turned

on per hour. The experiment lasted for 38 days, begin-

ning at 19:00PDT 2004-09-07 and ending at approximately

20:30PDT 2004-10-15.

Figure 3 shows a plot, similar to Figure 1, for the 69

Micron machines as measured by host2 but sub-sampled

to one out of every two packets. Note that the plot uses

different colors for the observed offsets for different ma-

chines (colors are overloaded). Since the slopes of the sets

of points for a machine corresponds to the machine’s skew,

this figure clearly shows that different machines in the lab

have measurably different clock skews. Thus, we can eas-

ily distinguish some devices by their clock skews (for other

devices, we cannot). Because Windows XP machines re-

set their TSopt clocks to zero when they reboot, some of

the diagonal lines seem to disappear several days into the

figure. Our algorithms handle reboots by recalibrating the

initial observed offset, though this recalibration is not vis-

ible in Figure 3. The time in Figure 3 begins on 8:30PDT

2004-09-10 (Friday) specifically because the administrators

of the lab tend to reboot machines around 8:00PDT, and

beginning the plot on Friday morning means that there are

fewer reboots in the figure. We consider this experiment in

more detail below, where our focus is on the stability of our

clock skew estimates.

STABILITY OF CLOCK SKEWS. We now consider the sta-

bility of the TSopt clock skews for the devices in the above-

mentioned undergraduate laboratory. Consider a single ma-

chine in the laboratory. We divide the trace for this machine

into 12- and 24-hour periods, discarding 12-hour periods

with less than 528 packets from the device, and discarding

24-hour periods with less than 1104 packets from the device

(doing so corresponds to discarding 12-hour periods when

the device is not up for at least approximately 11 hours, and

discarding 24-hour periods that the device is not up for at

least 23 hours). We compute the device’s clock skew for

each non-discarded period, and then compute the difference

between the maximum and minimum estimates for the non-

discarded periods. This value gives us an indication of the

stability of the device’s clock skew.

For 12-hour periods, the maximum difference for a sin-

gle device in the lab ranged between 1.29 ppm and 7.33

ppm, with a mean of 2.28 ppm. For 24-hour periods, the

maximum difference for a single device ranged between

0.01 ppm and 5.32 ppm, with a mean of 0.71 ppm. In-

terestingly, there seems to have been some administrator

function at 8:00PDT on 2004-09-10 that slightly adjusted

01224

36 48 60 72 84 96

time since start of measurement (hours)

-2

observed offset (seconds)

Figure 3. TSopt clock offset-sets for 69

Micron 448MHz Pentium II machines run-

ning Windows XP Professional SP1. Trace

recorded on host2, three hops away, 2004-

09-10 08:30PDT to 2004-09-14 08:30PDT.

the TSopt clock skews of some of the machines. If we con-

duct the same analysis for the trace beginning at 8:30PDT

2004-09-10 and ending on 2004-10-15,for 24-hour periods,

the range for maximum difference for each device in the lab

dropped to between 0.00 ppm and 4.05 ppm. See [15] for a

detailed table.

The current results strongly support our claim that mod-

ern processors have relatively stable clock skews. More-

over we believe that if the administrators of the lab allowed

us to exchange more packets with the 69 fingerprintees, we

would have found the clock skews to be even more stable.

In Section 6 we apply our clock skew estimates to a single

computer at multiple locations and on multiple dates, and

the skew estimates again are close (Table 3); our results be-

low further support our claim of the stability of clock skews

over time.

6 Access technology-, topology-, and

measurer-independent measurements

Here we consider our experiments which suggest that

clock skew estimates are relatively independent of the fin-

gerprintee’s access technology, the topology between the

fingerprintee and the measurer, and the measurer’s machine.

LAPTOPS IN MULTIPLE LOCATIONS. Our first set of ex-

periments along these lines measures laptop connected to

the Internet via multiple access technologies and locations

(Table 3). For all these experiments, laptop is a Dell Lat-

itude C810 notebook with a 1.133GHz Pentium III Mobile

processor and running a default installation of Red Hat 9.0

(Linux kernel 2.4.20-8). The measurer in all these experi-

Laptop location Start time (PDT) Duration Packets Wireless NAT Skew est.

San Diego, CA, home cable 2004-07-09, 22:00 3 hours 181 Ye s , W E P Yes −58.17 ppm

SD Supercomputer Center 2004-07-10, 10:00 3 hours 182 Yes No −58.00 ppm

CSE Dept, UCSD 2004-07-12, 12:00 3 hours 180 Yes No −58.24 ppm

San Diego, CA, home cable 2004-07-12, 21:00 3 hours 180 Yes Yes −58.21 ppm

Clinton, CT, home cable 2004-07-26, 06:00 3 hours 182 No Ye s −58.19 ppm

San Diego, CA, home cable 2004-09-14, 21:00 30 min 1795 Yes Ye s −58.22 ppm

SD Supercomputer Center 2004-09-22, 12:00 30 min 1765 Yes Ye s −58.13 ppm

San Diego dialup, 33.6kbps 2004-10-18, 10:00 30 min 1749 No No −57.57 ppm

SD Public Library 2004-10-18, 14:45 30 min 946 Yes Yes −57.63 ppm

Table 3. TCP timestamps-based skew estimates of laptop running Red Hat Linux 9.0 when connected

to host1 from multiple locations and when not running ntpd. The traces were recorded at host1.

ments, host1, is a Dell Precision 340 with a 2GHz Intel

Pentium 4 processor located within the UCSD Computer

Science and Engineering department and running Debian

3.0 with a recompiled 2.4.18 Linux kernel; host1 is also

configured to synchronize its system time with true time via

NTP.

For all experiments, we establish a TCP connection be-

tween laptop and host1, and then exchange TCP pack-

ets over that connection. On host1, we record a trace of

the connection using tcpdump. We then use our tech-

niques from Section 3 to estimate the skew of laptop’s

TSopt clock. As the horizontal line in Table 3 indicates, we

divide our experiments into two sets. In the first set, our ex-

periments last for three hours and exchange one TCP packet

every minute (we do this by performing a sleep(60) on

host1). For the second set of experiments, the connec-

tions last for 30 minutes, and a packet is exchanged at ran-

dom intervals between 0 and 2 seconds, as determined by a

usleep on host1. With few exceptions, the packets from

laptop are all ACKs with no data.

We conduct experiments when the laptop is connected

to the Internet via residential cable networks on both coasts

(Table 3). For our residential experiments,we use a 802.11b

wireless connection with 128-bit WEP encryption, a stan-

dard (unencrypted) 802.11b wireless connection, and a

standard 10Mbps 10baseT wired connection. We also con-

ducted experiments with our laptop connected to the San

Diego Supercomputer Center’s 802.11b wireless network,

from the UCSD Computer Science and Engineering wire-

less network, and from the San DiegoPublic Library’s wire-

less network. As the final column in the table shows, the

skew estimates are all within a fraction of a ppm of each

other. (If we subsample the first set of experiments to one

packet every 3 minutes, then the difference between the

skew estimates for any two measurements in the first set

is at most 0.45 ppm.)

PLANETLAB AND TOPOLOGY QUESTIONS. Although the

above results strongly suggest that skew estimates are inde-

pendent of access technology, the above experiments do not

stress-test the topology between the fingerprinter and the

fingerprintee. Therefore, we conducted the following set of

experiments. We selected a set of PlanetLab nodes from

around the world that reported, via ntptrace, approx-

imately accurate system times. We chose PlanetLab ma-

chines located at the University of California at San Diego,

the University of California at Berkeley, the University of

Washington, the University of Toronto (Canada), Prince-

ton (New Jersey), the Massachusetts Institute of Technol-

ogy, the University of Cambridge (UK), ETH (Switzerland),

IIT (India), and Equinix (Singapore). These PlanetLab ma-

chines, along with host1 and (in one case) CAIDA’s test

machine with a CDMA-synchronized Dag card, served as

our fingerprinters. Our fingerprintees were laptop and

host1,wherelaptop was connected both to the SDSC

wireless and to the CAIDA wired networks.

For each of our experiments, and for each of our cho-

sen PlanetLab nodes, we created a flow between the node

and the fingerprintee. Over each flow our fingerprintee sent

one packet at random intervals between 0 and 2 seconds;

here the fingerprintee executed usleep with appropriate

parameters. We then recorded the flows on the PlanetLab

machines using plabdump.Onhost1 we recorded the

corresponding flow using tcpdump. And on the machine

with the Dag card we used Coral [14] (that machine was

only reachable when laptop was connected directly to

CAIDA’s wired network). We then computed the skew us-

ing the techniques from Section 3. The results for laptop

are in Table 4. Notice that the skew estimates are in general

within a fraction of a ppm of each other, suggesting that our

skew estimates are independent of topology.

For distance measurements for Table 4, we used tracer-

oute to determine hop count, and then used mean time be-

tween when tcpdump recorded a packet on the measured

device and the time between when plabdump recorded the

packet on the measurer. This distance estimate also includes

laptop, 2004-09-17, 08:00–10:00 PDT laptop, 2004-10-08, 08:00–10:00 PDT

Measurer Skew estimate Dist. from measurer Skew estimate Dist. from measurer

host1 −58.23 ppm 7 hops, 2.77 ms −58.03 ppm 8 hops, 1.16 ms

San Diego, CA −58.07 ppm 7 hops, 1.21 ms −58.03 ppm 8 hops, 1.15 ms

Berkeley, CA −58.17 ppm 10 hops, 4.02 ms −58.02 ppm 12 hops, 5.06 ms

Seattle, WA −58.15 ppm 8 hops, 14.74 ms −58.01 ppm 9 hops, 15.12 ms

Toronto, Canada −58.31 ppm 16 hops, 44.43 ms

Princeton, NJ −57.97 ppm 13 hops, 37.59 ms −57.91 ppm 14 hops, 36.97 ms

Boston, MA −57.93 ppm 12 hops, 35.82 ms −58.10 ppm 13 hops, 41.09 ms

Cambridge, UK −58.06 ppm 20 hops, 84.19 ms −58.18 ppm 21 hops, 86.45 ms

ETH, Switzerland −58.38 ppm 20 hops, 90.51 ms −58.40 ppm 21 hops, 84.07 ms

IIT, India −59.60 ppm 16 hops, 199.27 ms

Equinix, Singapore −58.10 ppm 18 hops, 99.50 ms −58.05 ppm 15 hops, 93.55 ms

CAIDA test lab −57.98 ppm 5 hops, 0.24 ms

Table 4. Skew estimates of laptop, running Red Hat 9.0 with ntpd, for traces taken simultaneously

at multiple locations. On 2004-09-17 the laptop was connected to the SDSC wireless network, and on

2004-10-08 the laptop was connected to the CAIDA wired network. The Toronto and India lines have

empty cells because the PlanetLab machines at those locations were down during the experiment.

The Boston machine on 2004-10-08 was a different PlanetLab machine than the one on 2004-09-17.

The empty cell for the CAIDA test lab is because the lab is only reachable from CAIDA’s wired network.

the time spent in the application layers on the machines, but

should give a rough estimate of the time it takes packets to

go from the fingerprintee to the measurer.

The results of these experiments suggest that our TSopt

clock skew estimation technique is generally independent

of the topology and distance between the fingerprinter and

the fingerprintee. Furthermore, these results suggest that

our skew estimation technique is independent of the actual

fingerprinter, assuming that the fingerprinter synchronizes

its system time with NTP [19] or something better [26].

7 Effects of operating system, NTP, and spe-

cial cases

OPERATING SYSTEMS AND NTP ON FINGERPRINTEE.In

Table 5 we show skew estimates for the same physical de-

vice, laptop, running both Red Hat 9.0 and Windows XP

SP2, and both with and without NTP-based system clock

synchronization. (For this experiment, laptop sent one

packet to the measurer, host1, at random intervals be-

tween 0 and 2 seconds; laptop was connected to the

SDSC wireless network, and was 7 hopsaway from host1;

host1 also sent a ICMP Timestamp Request to laptop

at random intervals between 0 and 60 seconds.) The ta-

ble shows that, for the listed operating systems, the system

clock and the TSopt clock effectively have the same clock

skew when the device’s system time is not synchronized

with NTP, and that the TSopt clock skew is independent of

whether the device’s system clock is maintained via NTP.

Although not shown in the figure, our experiments with

OpenBSD 3.5 on another machine suggest that the TSopt

clock and system clock on default OpenBSD 3.5 installa-

tions have the same skew (approximately 68 ppm). On the

other hand, at least with this test machine, the TSopt clock

and system clock on a default FreeBSD 5.2.1 system have

different skews (the TSopt clock skew estimate is about the

same as with OpenBSD, but the system clock skew estimate

is approximately 80 ppm). When we turn on ntpd under

FreeBSD 5.2.1, the TSopt clock skew remained unchanged.

POWER OPTIONS FOR LAPTOPS. We also consider how

the clock skews of devices are affected by the power op-

tions of laptops. In the case of Red Hat 9.0, when laptop

is running with the power connected, if we switch to battery

power, there is a brief jump in the TSopt clock offset-set for

the device, and then the device continues to have the same

(within a fraction of a ppm) clock skew. For laptop run-

ning Windows XP SP2, if the laptop is idle from user input

but continues to maintain a TCP flow that we can monitor,

then the TSopt clock skew changes after we switch to bat-

tery power. If we repeat this experiment several times, and

if we boot with only battery power, we find that the clock

skews with battery power are in all cases similar. When

booting with outlet power, the clock skew on laptop run-

ning Windows XP initially begins with a large magnitude,

and then stabilizes to a skew like that in Table 5 until we

disconnect the power; the initially large skew may be due

to the laptop recharging its batteries. We have not sam-

pled a large enough set of laptops to determine whether the

Start time Operating System NTP skew estimate skew estimate

(TCP tstamps) (ICMP tstamps)

2004-09-22, 12:00 PDT Red Hat 9.0 No −58.20 ppm −58.16 ppm

2004-09-17, 08:00 PDT Red Hat 9.0 Ye s −58.16 ppm −0.14 ppm

2004-09-22, 21:00 PDT Windows XP SP2 No −85.20 ppm −85.42 ppm

2004-09-23, 21:00 PDT Windows XP SP2 Yes −84.54 ppm 1.69 ppm

Table 5. Experiments for the same physical device, laptop, running different operating systems and

with NTP synchronization both on and off. For all experiments, laptop was located on the SDSC

wireless network. Additionally, laptop was up for an hour before the Windows measurements.

14700 15000 15300 15600

time since start of measurement (seconds)

-300

-200

-100

100

200

300

observed offset (ms)

Figure 4. TSopt clock offset-sets for 100

honeyd 0.8b Windows XP SP1 virtual hosts.

Start time: 2004-09-19, 23:00PDT; honeyd

running on host3. Points are connected in

this figure to highlight the correlation be-

tween the virtual hosts.

clock skews with battery power are a simple function of the

clock skews with outlet power, though the skews with bat-

tery power seem to be consistent for a single laptop.

8 Applications

We now consider some applications of our techniques,

though we emphasize that our most important results are

the foundations we introduced in the previous sections that

make the following applications possible.

VIRTUALIZATION AND VIRTUAL HONEYNETS.Wecre-

ated a honeyd [24] version 0.8b virtual honeynet consist-

ing of 100 Linux 2.4.18 virtual hosts and 100 Windows XP

SP1 virtual hosts. Our server in this experiment, host3,

is identical to host1, has 1GB of RAM, and maintains its

system time via NTP. We ran honeyd with standard nmap

and xprobe2 configuration files as input; honeyd used

the information in these files to mimic real Linux and Win-

dows machines. We ran nmap and xprobe2 against the

virtual hosts to verify that nmap and xprobe2 could not

distinguish the virtual hosts from real machines.

We applied our TCP timestamps- and ICMP-based skew

estimation techniques to all 200 virtual hosts. Our fin-

gerprinter in this experiment was on the same local net-

work. We observed several methods for easily distinguish-

ing between honeyd virtual hosts and real machines. First,

we noticed that unlike real Linux and Windows machines,

the virtual hosts always returned ICMP Timestamp Replies

with zero in the transmit timestamp field. Additionally, we

observed that the honeyd Windows XP virtual hosts had

TSopt clocks Ctcp with Hz[Ctcp]=2, whereas all of the real

Windows XP machines that we tested had Hz[Ctcp ]=10.

The lesson here is that although the nmap and xprobe2

configuration files provide enough information for the re-

spective programs to effectively fingerprint real operating

systems, the configuration files do not provide enough in-

formation for honeyd to be able to correctly mimic all as-

pects of the Linux and Windows protocol stacks.

Even if honeyd completely mimicked the network

stacks of real Linux 2.4.18 and Windows XP SP1 machines,

we could still use our remote physical device fingerprint-

ing techniques to distinguish between our 200 virtual hosts

and 200 real machines. Our TSopt clock skew estimates for

all 200 virtual hosts were approximately zero and the sys-

tem clock skew estimates for all 200 virtual hosts were ap-

proximately the same positive value. Given our discussion

in Section 5 of the distribution of clock skews, this lack

of variability in clock skews between virtual hosts is not

what one would expect from real machines. Furthermore,

the TSopt and system clocks between all the virtual hosts

of the same operating system were highly correlated; e.g.,

Figure 4 shows the TSopt offset-sets for all 100 Windows

XP SP1 virtual hosts 241 minutes into our experiment. We

communicated our results to the author of honeyd and, in

response, version 1.0 of honeyd randomly assigns TSopt

clock skews to each virtual host using a Gaussian distribu-

tion around the server’s system time. This decision may

affect other components of the system, e.g., if the server

runs ntpd, changes to the server’s system time may appear

as global changes to the distribution of the virtual hosts’

clocks. Version 1.0 of honeyd still issues ICMP Times-

tamp Replies with zero transmit timestamps. Furthermore,

the system clocks on version 1.0 honeyd virtual hosts are

still highly synchronized and are too fast by several orders

of magnitude.

To experiment with real virtualization technologies, we

installed VMware Workstation 4.5.2 on host3, but this

time host3 ran Red Hat 9.0. We then installed five default

copies of Red Hat 9.0 under VMware. We applied our skew

estimation techniques to these five virtual machines, as well

as to host3. The results show that the five virtual machines

do not have constant (or near constant) clock skews, shown

by the non-linearity of the points in Figure 5. Furthermore,

the magnitude of the clock skews on these virtual machines

is larger than we would expect for physical machines. We

feel confident that these observations and natural extensions

could prove useful in distinguishing virtual honeynets from

real networks.

COUNTING THE NUMBER OF DEVICES BEHIND A NAT.

Another natural application of our techniques is to count the

number of devices behind a NAT. To briefly recall previous

work in this area, Bellovin [7] showed that an adversary

can exploit the IP ID field to count the number of devices

behind a NAT, but his approach is limited in three ways:

(1) the IP ID field is only 16-bits long; (2) recent operating

systems now use constant or random IP ID fields; and (3)

his technique cannot count the total number of devices be-

hind a NAT if not all of them are active at the same time.

Our suggested approach to this problem has two phases.

First, partition the trace into (candidate) sets corresponding

to different sequences of time-dependent TCP timestamps;

creating such a partition is relatively easy to do unless two

machines have approximately the same TSopt clock values

at some point in time, perhaps because the machines booted

at approximately the same time. Then apply our clock skew

estimation techniques to each partition, counting hosts as

unique if they have measurably different clock skews. If two

devices have approximately the same TSopt clock values

at some point in time but have measurably different clock

skews, then one can detect and correct this situation in the

analysis of the partition’s offset-set.

FORENSICS AND TRACKING INDIVIDUAL DEVICES.The

utility of our techniques for forensics purposes follows

closely from our claims (1) that there is variability in the

clock skews between different physical devices (Section 5),

(2) that the clock skew for a single device is approximately

constant over time (Section 5), and (3) that our clock skew

estimates are independent of access technology, topology,

and the measurer (Section 6). For forensics, we anticipate

that our techniques will be most useful when arguing that a

0600 1200 1800 2400 3000 3600 4200 4800 5400 6000 6600 7200

time since start of measurement (seconds)

-4

-3

-2

-1

observed offset (seconds)

Figure 5. TSopt clock offset-sets for five

VMware Workstation virtual machines run-

ning Red Hat 9.0, and for the host, host3,

also running Red Hat 9.0. 2004-10-27 17:00–

19:00PDT. The top set of points corresponds

to the TSopt clock offset set for host3.

given device was not involved in a recorded event. With re-

spect to tracking individual devices, we stress that our tech-

niques do not provide unique serial numbers for devices,

but that our skew estimates do provide valuable bits of in-

formation that, when combined with other sources of infor-

mation such as operating system fingerprinting results, can

help track individual devices on the Internet.

UNANONYMIZING ANONYMIZED DATA SETS.Itiscom-

mon for organizations that provide network traces contain-

ing payload data to anonymize the IP addresses in the traces

using some prefix-preserving anonymization method [28,

29]. If an organization makes available both anonymized

and unanonymized traces from the same link, one can

use our techniques to catalyze the unanonymization of the

anonymized traces. Such a situation is not hypothetical: in

addition to the 2004-04-28 trace that we used in Section 5,

CAIDA took another trace from the same link on 2004-04-

21, but the 2004-04-21 trace included payload data and was

therefore anonymized.

To study how one might use our clock skew estima-

tion techniques to help unanonymize anonymized traces,

on 2005-01-13 and 2005-01-21 CAIDA took two two-hour

traces from a major OC-48 link (the same link from which

CAIDA captured the 2004-04-28 trace). We anonymized

the 2005-01-13 trace and experimented with our ability to

subsequently unanonymize it. Given the value of a de-

vice’s TSopt clock and knowledge of that clock’s intended

frequency Hz, we can compute the approximate uptime of

the device. (Prior to our work, one method for inferring

Hz from a passive trace would be to use a program like

p0f [3].) As a first attempt at unanonymizing the 2005-01-

13 trace, we paired anonymized IP addresses from 2005-

01-13 with IP addresses from 2005-01-21 when our uptime

estimate of a host in 2005-01-21 is eight days higher (plus

or minus five minutes) than the uptime of a host in 2005-01-

13 and when both hosts have the same TTLs and intended

frequencies. Our program produced 4613 pairs of candidate

anonymous to real mappings, of which 2660 (57.66%) were

correct. To reduce the number of false matches, especially

for small uptimes, we modified our program to filter out

pairs that have TSopt clock skews different by more than 3

ppm. We also incorporated our clock skew estimates into

our uptime estimates. These changes reduced the number

of candidate mappings to 2170, of which 1902 (87.65%)

were correct. There are a total of 11862 IP addresses in

both the 2005-01-13 and 2005-01-21 traces that have the

TCP timestamps option enabled. Since the anonymiza-

tion is prefix-preserving, given the candidate mappings one

can begin to unanonymize address blocks. We are un-

aware of any previous discussion of the problems to prefix-

preserving anonymization caused by leaking information

about a source via the TCP timestamps option.

9 Other measurement techniques

Although the techniques we describe above will likely

remain applicable to current generation systems, we suspect

that future generation security systems might try to resist

some of the physical device fingerprinting techniques that

we uncover. In anticipation of these future systems, we con-

sider possible avenues for clock-based physical device fin-

gerprinting when information about a system’s TSopt clock

or system clock is not readily available to an adversary; we

do not consider here but recognize the possibility of finger-

printing techniques that profile other aspects of a device’s

hardware, e.g., processor speed or memory. These direc-

tions assume that new operating systems mask or do not in-

clude the TSopt clock values in the TCP headers and do not

reply to ICMP Timestamp Requests, but that the systems’

underlying clocks still have non-negligible skews. (This as-

sumption may not be valid if, for example, at boot a new op-

erating system does a more precise estimation of the oscilla-

tor frequencies supplying the hardware basis for the clocks.)

The techniques we propose in this section are less refined

than the techniques elsewhere in this paper; we envision

them as starting points for more sophisticated techniques.

FOURIER TRANSFORM. Some systems send packet at 10

or 100 ms intervals, perhaps due to interrupt processing

or other internal operating system feature on one side of

a flow. When this condition holds, we can use the Fourier

transform to extract information about the system’s clock

skew. Figure 6 plots the TSopt clock offset-sets for a de-

vice in BBSwith a 2 Hz TSopt clock. The five diagonal

0 900 1800 2700 3600 4500 5400 6300 7200

time since start of measurment (seconds)

500

1000

1500

2000

offset set (ms)

TSopt clock skew estimate by linear programming upper bound

Figure 6. TSopt clock skew estimate for a

source in BBS. Trace recorded on an OC-48

link of a U.S. Tier 1 ISP, 2004-04-28 19:30–

21:30PDT. TSopt clock skew estimate via lin-

ear programming: 175.2 ppm. Clock skew es-

timate via the Fourier transform: 175.6 ppm.

bands suggests that the machine clusters packet transmis-

sions at approximately 100 ms intervals, and we can use the

Fourier transform on packet arrival times to estimate the fre-

quency at which the device actually transmits packets (here

packet arrival times refers to the times at which the moni-

tor records the packets). For the source shown in Figure 6,

after computing the Fourier transform, the frequency with

the highest amplitude was 25.00439, which implies a skew

of 25.00439/25 −1,or175.6ppm. Moreover the top 19

frequencies output by the Fourier transform all imply skews

between 171.0ppm and 179.3ppm. These values are all

close to the 175.2ppm output by our TCP timestamps-based

approach but do not make any use the TCP timestamps con-

tained with the packets.

Although our Fourier-based technique does not require

knowledge of a device’s TSopt or system clocks, our

Fourier-based solution is currently not automated. This lack

of automation, coupled with the fact that current generation

systems readily relinquish information about their TSopt

and system clocks, means that our Fourier-based solution

is currently less attractive than the techniques we described

in Sections 3 and 4.

PERIODIC USER-LEVEL ACTIVITIES. Toward estimating

the system clock skew of devices that do not synchronize

their system times with NTP, we note that many applica-

tions perform certain operations at semi-regular intervals.

For example, one can configure most mail clients to poll for

new mail every nminutes. As another example, Broido,

Nemeth, and claffy show that some Microsoft Windows

2000 and XP systems access DNS servers at regular inter-

vals [8]. It may be possible to infer information about a

device’s system clock skew by comparing differences be-

tween actual intervals of time between these periodic activ-

ities and what the application intends for those intervals of

time to be.

10 Conclusions

In this study we verified the ability and developed tech-

niques for remote physical device fingerprinting that exploit

the fact that modern computer chips have small yet non-

trivial and remotely detectable clock skews. We showed

how our techniques apply to a number of different practi-

cally useful goals, ranging from remotely distinguishing be-

tween virtual honeynets and real networks to counting the

number of hosts behind a NAT. Although the techniques we

described will likely remain applicable to current generation

systems, we suspect that future generation security systems

might offer countermeasures to resist some of the finger-

printing techniques that we uncover. In anticipation of such

developments, we discussed possible avenues for physical

device fingerprinting when information about a system’s

TSopt clock or system clock are not readily available to the

adversary. Our results compellingly illustrate a fundamental

reason why securing real-world systems is so genuinely dif-

ficult: it is possible to extract security-relevant signals from

data canonically considered to be noise. This aspect renders

perfect security elusive, and even more ominously suggests

that there remain fundamental properties of networks that

we have yet to integrate into our security models.

Acknowledgments

We thank Bruce Potter and Stefan Savage for helpful dis-

cussions, Emile Aben, Dan Andersen, Colleen Shannon,

and Brendan White for collecting some of the traces that

we analyzed, and William Griswold for loaning us a PDA

from the HP Mobile Technology Solutions gift to the Ac-

tiveCampus project. All three authors were supported by

the SciDAC program of the US DOE (award # DE-FC02-

01ER25466). T. Kohno was also supported by NDSEG and

IBM Ph.D. Fellowships.

References

[1] Endace measurement systems, 2004. URL: http://www.

endace.com/.

[2] Nmap free security scanner, 2004. URL: http://www.

insecure.org/nmap/.

[3] Project details for p0f, 2004. URL: http://

freshmeat.net/projects/p0f/.

[4] VMware virtual infrastructure, 2004. URL: http://www.

vmware.com/.

[5] Xprobe official home, 2004. URL: http://www.

sys-security.com/html/projects/X.html.

[6] K. G. Anagnostakis, M. Geenwald, and R. S. Ryger. cing:

Measuring network-internal delays using only existing in-

frastructure. In INFOCOM, 2003.

[7] S. M. Bellovin. A technique for counting NATted hosts. In

IMW, 2002.

[8] A. Broido, E. Nemeth, and k. claffy. Spectroscopy of DNS

update traffic. In SIGMETRICS, 2003.

[9] S. Donnelly. High precision timing in passive measure-

ments of data networks. Ph.D. thesis, University of Waikato,

Hamilton, New Zealand, 2002.

[10] M. E. Dyer. Linear time algorithms for two- and three-

variable linear programs. SIAM J. Comput., 13, 1984.

[11] I. D. Graham, M. Pearson, J. Martens, and S. Donnelly. Dag

- A cell capture board for ATM measurement systems. URL:

http://dag.cs.waikato.ac.nz/.

[12] R. L. Graham. An efficient algorithm for determining the

convex hull of a finite planar set. Inf. Process. Lett., 1, 1972.

[13] V. Jacobson, R. Braden, and D. Borman. TCP extensions for

high performance. RFC 1323, May 1992.

[14] K. Keys, D. Moore, R. Koga, E. Lagache, M. Tesch, and

k. claffy. The architecture of the CoralReef Internet traffic

monitoring software suite. In PA M , 2001.

[15] T. Kohno, A. Broido, and k. claffy. Remote physi-

cal device fingerprinting. Full version of this paper,

available at http://www-cse.ucsd.edu/users/

tkohno/papers/PDF/, 2005.

[16] N. Megiddo. Linear-time algorithms for linear programming

in r3and related problems. SIAM J. Comput., 12, 1983.

[17] J. Micheel, S. Donnelly, and I. Graham. Precision times-

tamping of network packets. In IMW, 2001.

[18] D. Mills. Simple network time protocol (SNTP) version 4

for IPv4, IPv6 and OSI. RFC 2030, 1996.

[19] D. L. Mills. Network time protocol (version 3): Specifica-

tion, implementation and analysis. RFC 1305, 1992.

[20] S. B. Moon, P. Skelly, and D. Towsley. Estimation and re-

moval of clock skew from network delay measurements. In

INFOCOM, 1999.

[21] A. P´asztor and D. Veitch. PC based precision timing without

GPS. In SIGMETRICS, 2002.

[22] V. Paxson. On calibrating measurements of packet transit

times. In SIGMETRICS, 1998.

[23] J. Postel. Internet control message protocol. RFC 792, 1981.

[24] N. Provos. A virtual honeypot framework. In Usenix Secu-

rity 2004, 2004.

[25] C. Shannon. The mathematical theory of communication.

1949. Urbana, University of Illinois Press.

[26] D. Veitch, S. Babu, and A. P´asztor. Robust synchronization

of software clocks across the Internet. In IMC, 2004.

[27] F. Veysset, O. Courtay, and O. Heen. New tool

and technique for remote operating system fingerprint-

ing, 2002. URL: http://www.intranode.com/fr/

doc/ring-short-paper.pdf.

[28] J. Xu, J. Fan, M. Ammar, and S. B. Moon. On the design and

performance of prefix-preserving IP traffic trace anonymiza-

tion. In IMW, 2001.

[29] J. Xu, J. Fan, M. H. Ammar, and S. B. Moon. Prefix-

preserving IP address anonymization: Measurement-based

security evaluation and a new cryptography-based scheme.

In ICNP, 2002.