remote physical device fingerprinting authors tadayoshi kohno, andre broido, kc claffy appears in...

Remote Physical Device FingerprintingAuthorsTadayoshi Kohno, Andre Broido, KC Claffy

Appears inIEEE Symposium on Security and Privacy, 2005

Presented byPeter Matthews

Introduction

There are a number of reliable techniques for remote operating system fingerprintingnmapXProbe

Paper proposes the next step:Remotely fingerprint a physical device without

that device's knowledge or cooperation

Clock Drift

A standard clock circuit in a computer system uses a quartz crystal oscillator as its time base, similar to any modern wristwatchSome amount of imprecision in the oscillatory

frequency Clocks using these thus exhibit drift over

time when compared to the actual time

Clock Skew A clock will usually have some offset

Offset(t) = time_reported(t) – true_time(t) The clock skew S is the change in this

offset over timeS = d Offset(t) / dt

Measured in ppm (μs/s)

How much skew? +/- 4 seconds a day common (25 minutes a year)

Importantly, paper argues skew of a device is (generally) consistent and distinctive to that device Thus can use as a fingerprint for this device

24 hours later

Network Time Protocol (NTP) Hierarchical server system

Top-level servers synchronized via atomic clocks, GPS

Used to synchronizes system clock periodically

Techniques for Determining Clock Skew of a Remote Device 3 approaches given

TCP Timestamps Option (TSopt)

32-bit timestamp contained in each packet, taken from a virtual clock that is “at least approximately proportional to real time”

Frequency (Hz) between 1 and 1000 Windows 2000, XP – 10 Hz (100 ms resolution) Redhat 9.0 – 100 Hz (10 ms resolution)

Usually reset to zero upon reboot Usually not affected by changes to the device's

system clock (NTP synchronization does not affect)

Exploiting the TCP Timestamps Option (Passive Approach) The measurer – any entity capable of observing

TCP packets from the fingerprintee Create a trace of TCP packets from fingerprintee For each packet plot a point

X value: Amount of actual time passed between reception of first packet in trace and the current packet

Y value: The offset observed for this packet, based on timestamp

Use linear programming to determine the equation of the line y = αx + β that best upper-bounds this set of pointsα is the estimate of the clock skewβ is an initial observed offset

TSopt clock skew estimates for two sources from a OC-48 link of a US Tier 1 ISP over a two hour period.

Exploiting the TCP Timestamps Option (Semi-Passive Approach) Windows 2000 and XP machines do not

set timestamp flag in their initial SYN packets

RFC 1323 mandates that none of the following TCP packets in the connection can include timestamp

Thus, previous approach will not work if a Windows machine is behind NAT, firewall

Paper’s trick: The measurer includes timestamp in the responding SYN/ACK packet

Windows machines then include timestamp in all subsequent packets of this connection

SYN, TSopt=0SYN, TSopt=1

SYN-ACK, TSopt=1

data, timestamp data, timestamp

Subsequently…

ICMP Timestamps

Reports value of system clock (milliseconds past midnight)

RFC 792 requires frequency is 1000 Hz (1 ms resolution)

If system clock is updated via NTP regularly, will be relatively accurateHowever, most hosts do so infrequently

Exloiting ICMP Timestamp Requests (Active Approach) The measurer: entity capable of sending

ICMP Timestamp Request and storing the fingerprintee's subsequent ICMP Timestamp Reply messages

Limitation: Fingerprintee must not be behind a NAT or firewall that filters ICMP

Estimation of clock skew is similar to that in TSopt methods.

Questions Concerning Clock Skew

What is the distribution of clock skews among devices?

How stable are these clock skews over time?

Can these clock skews be measured accurately, independent of network topology and access technology?

Distribution of Clock Skews

Figure 1: Histogram of TSopt clock skew estimates for sources in a 2 hour network trace from a OC-48 link of a US Tier 1 ISP. (Considered only sources that sent packet over a period of at least 50 minutes per hour, and sent at least 2000 packets per hour.)

Could this skew simply reflect different operating system and hardware configurations?

To answer this, TSopt clock offsets were measured for 69 Micron 448 MHz Pentium II machines running Windows XP SP1 over 38 days

~48 TCP packets with timestamp per hour

TSopt clock offsets measured for 69 homogenous machines over 96 hours.

Clock skew can be used to distinguish some, but not all machines.

Stability of Clock Skews The network traces from this experiment were

then divided into 12 and 24 hour periods All periods of the same length for each machine were

then compared Range of maximum difference for a single

device: 12-hour period: 1.29 – 7.33 ppm, 2.28 ppm mean 24-hour period: 0.01 – 5.32 ppm, 0.71 ppm mean

Supports authors’ claim that modern processors have relatively stable clock skews

Independence of Access Technology

Fingerprinting of a laptop connected via various access technologies and locations

Skew estimates all within a 1 ppm range

Independence of Network Topology

PlanetLab machines located globally used as measurer, same laptop as previous experiment

Excepting measurement from IIT, skew estimates within .5 ppm range

Generally speaking, estimates are largely independent of topology and distance between fingerprinter and fingerprintee

Other OS Factors

For tested operating systems, system clock and TSopt clock effectively have the same clock skew

Applications

Detecting virtual honeynets and virtual hosts Counting number of devices behind a NAT Tracking individual devices

With some probability

Arguing that a given device was not involved in a recorded event

Could use un-anonymized traces on a link to assign hosts in an anonymized trace

Strengths

Shows that it is possible to extract relevant security information from data considered noise

Approach could be used with any other protocols that leak information about a device’s clock

Weaknesses

Further experimentation required Laptop running Windows XP SP2 has a noticeably

different TSopt clock skew after switching to battery power

Newer processors throttle their speeds based on temperature and load, affects voltage from power supply

Easy to circumvent particular methods echo 0 > /proc/sys/net/ipv4/tcp_timestamps Randomize TSopt timestamp Filter ICMP timestamp

Possible Extensions of Paper’s Approach Utilization of approach with other protocols

that leak information about a device’s clock

Use of profiling in combination with skew dataSkew is within a certain range and machine

visits certain websites frequentlyOS profiling techniques

Questions?