DOF: A Local Wireless Information Plane

molassesitalianΤεχνίτη Νοημοσύνη και Ρομποτική

6 Νοε 2013 (πριν από 3 χρόνια και 5 μήνες)

54 εμφανίσεις

DOF:


A Local Wireless Information Plane

Stanford University

Steven Hong

Sachin Katti


1

August 17, 2011

Problem

Unlicensed spectrum (e.g. ISM Band
-

2.4 GHz) has historically been
managed “socially


How can we design a smart radio which
maximizes throughput while causing minimal
harm to coexisting radios?

2

Can we use current mechanisms to design
these smart radios?

Current coexistence mechanisms


Carrier Sense, RTS/CTS


Rate Adaptation


Adaptive Frequency
Hopping


...

Current mechanisms are not sufficient for
designing high performance smart radios

3

How would we build a smart radio
which coexists with legacy devices?

Microwave

Smart
Transmitter

Smart Receiver

1.
The protocol types operating in the local vicinity

2.

The spectrum occupancy of each type

3.

The spatial directions of each type

Knowledge of

WiFi AP

Heart Monitor

AoA

Freq

2.3
GHz

2.5
GHz

0
°

180
°

AoA

Freq

2.3
GHz

2.5
GHz

0
°

180
°

Freq

2.3
GHz

2.5
GHz

Freq

2.3
GHz

2.5
GHz

4

DOF

(
D
egrees
O
f
F
reedom)

Local wireless information plane which provides all 3
of these quantities (type, spectral occupancy, spatial
directions) in a single framework

DOF Performance
Summary


DOF is
robust to SNR of detected
signals


Accurate at received signals as low as 0dB



DOF is
robust to
multiple overlapping signals


Accurate even when three unknown signals are present



DOF is relatively computationally inexpensive


Requires 30% more computation over standard FFT

5

DOF: High Level Architecture

Feature

Extraction

{
𝒑
,
𝑭

(
𝒊
)

}
𝒏
=

𝑵

DOF Estimation

(AoA Detection)

ADC

Signal

Time Samples

Classification

F( i )
MAC

DOF

{
𝒑
,
𝑭
𝒄
,
𝑩𝑾
}
𝒏
=

𝑵

{
𝚯
}
𝒏
=

𝑵

{
𝚯
,
𝒑
,
𝑭
𝒄
,
𝑩𝑾
}
𝒏
=

𝑵

DOF Estimation

(Spectrum Occupancy)

6

DOF operates on windows of raw
time samples from the ADC

Raw samples are processed to extract
feature vectors

Feature Vectors are used to detect

1.
Signal Type

2.
Spectral Occupancy

3.
Spatial Directions

The MAC layer utilizes this mechanism
to inform its coexistence policy

For almost all “man
-
made” signals


there are hidden repeating
patterns that are unique and necessary for operation

Key Insight

CP

CP

CP

Data

Data

Data

…………………….

Repeating Patterns in WiFi OFDM signals

Repeating Patterns in Zigbee signals

Time

Leverage unique patterns to infer 1) type, 2)
spectral occupancy, and 3) spatial directions

7

Pattern Frequency
(
α
)

Delay (
τ
)

Advantages


Robustness to noise,


Uniqueness for each
protocol

Extracting Features from Patterns

8

If a signal has a repeating pattern, then when we


Correlate the received signal against itself delayed by a fixed amount, the
correlation will peak when
the delay is equal to the period at which the
pattern repeats.



𝑥
𝛼
𝜏
=

𝑥

𝑥



𝜏


𝑗2𝜋𝛼𝑛

𝑛

Pattern Frequency (
𝛼
)


The frequency at which the pattern repeats

Disadvantage
: Computationally expensive to
calculate the patterns in this manner

Cyclic Autocorrelation Function (CAF)

Feature Extraction:

Efficient Computation

The CAF can be represented using an equivalent form called the Spectral
Correlation Function (SCF)


SCF can be calculated
for Discrete Time
Windows using just
FFTs


9


𝑥
𝛼

=


𝑥
𝛼
𝜏


𝑗2𝜋𝑓𝜏

𝜏
=



=





1
𝐿

𝑋
𝑙𝑁

𝑋
𝑙𝑁

(


𝛼
)

𝐿

1
𝑙
=
0

Pattern Frequency
(
α
)

Frequency (f)

WiFi Spectral Correlation Function

Feature Vectors
are calculated by
computing


𝜶


at different values of
𝜶

Classifying Signal Type

10


Single signals are well separated in the feature vector space,
𝐹


Works well when there is a single signal but fails
when there are multiple interfering signals

Feature Dimension 1:





𝜶




Feature Dimension 2:


𝜶




Support Vector Machines (SVM) can be used to classify signal type,


Multiple interfering signals are not
straightforward to classify


Multiple signals are made up of components and features of single
signals, making them difficult to distinguish

Need a robust algorithm to determine the
number of interfering signals

Feature Dimension 1:





𝜶




Feature Dimension 2:


𝜶



Feature Dimension 1:





𝜶




Feature Dimension 2:


𝜶



11

1) Real signal packets are asynchronous

12

Inferring the number of signals:

Exploiting Asynchrony

ZigBee

t

Overlapping Packets

WiFi

Nonzero Components in

F( i )
Received Signal

2
) This asynchrony shows up in as an increase or decrease in the



number of non
-
zero components

F( i )
Measuring differences in
𝑭
𝒊

is more robust
than differences in energy

DOF: High Level Architecture

Feature

Extraction

{
𝒑
,
𝑭

(
𝒊
)

}
𝒏
=

𝑵

DOF Estimation

(AoA Detection)

ADC

Signal

Time Samples

Classification

F( i )
MAC

Asynchrony
Detector/

Power
Normalization

SVM
-
1

SVM
-
N

Counter++

. . .

If
Δ
L0>Threshold

Counter
--

If
Δ
L0<
-
Threshold

Sig1 Class

Sig
i

Class

1 Signal

N
Signals

SigN Class

. . .

. . .

While

DOF = Active

DOF

{
𝒑
,
𝑭
𝒄
,
𝑩𝑾
}
𝒏
=

𝑵

{
𝚯
}
𝒏
=

𝑵

{
𝚯
,
𝒑
,
𝑭
𝒄
,
𝑩𝑾
}
𝒏
=

𝑵

The signal types can be leveraged along with
the feature vectors to estimate




1) Spectrum Occupancy


2) Spatial Directions

DOF Estimation

(Spectrum Occupancy)

13

Estimating Spectrum Occupancy

14


Communication signals are sequences of periodic pulses



=

2𝜋




𝑗2𝜋
𝑓
𝑐
𝑡



These pulses are patterns embedded within the signal which repeat at
a particular frequency


These frequencies at which these patterns repeat tell us the
bandwidth



and carrier frequency



of the signal

1

0

Bit Sequence

b

Amplitude modulated Pulse


2𝜋




Pulse multiplied by Carrier Wave


2𝜋




𝑗2𝜋
𝑓
𝑐
𝑡

Estimating Spectrum Occupancy

15

Modulated Zigbee Signal

Time


Because these patterns repeat, they are natural components of the
feature vector

Relationship between feature vector and Bandwidth/Carrier Frequencies

Signal Type

Feature Vector Frequencies

WiFi

Bluetooth

ZigBee

all
𝜶

𝒔

between
[

𝒄

𝑩𝑾

,


𝒄
+
𝑩𝑾

]



𝒄
,


𝒄

𝑩𝑾

,



𝒄
+
𝑩𝑾





𝒄
+
𝑩𝑾
,



𝒄
+
𝑩𝑾

DOF leverages this relationship to compute the
spectral occupancy of each signal type

Pattern Frequency
(
α
)

Frequency (f)

ZigBee Spectral Correlation Function

Estimating Angles of Arrival

. . .

1

2

M

Incoming Signal

d

Array Elements


Each array
element experiences a

delay
of
τ

relative to the first
array
element, which is a function of the
Angle of Arrival (
AoA
)

. . .

16


This unique delay induces a particular characteristic on the
feature vector
𝐹

(
𝑖
)

which can be computed

DOF uses the same feature vector to infer
1)type, 2)spectral occupancy, 3)spatial directions

Implementation

R
X
2

R
X

1

R
X
3


Channel t
races
were collected using a
modified channel sounder with
a frontend bandwidth of 100MHz spanning the entire ISM band.



Wideband Radio Receiver placed at 3 different locations while
transmitter was placed randomly in the office



Raw Digital Samples are collected and processed offline on a PC with
Intel Core i7 980x Processor and 8GB RAM

17

Compared Approaches

Identifying
Protocol
Types


RF
Dump (CoNEXT 2009)


Energy Detection + Packet
Timing

Estimating Spectrum Occupancy


Jello (NSDI 2010)


Edge Detection on Power Spectral
Density

Estimating Angles of Arrival


Secure Angle (HOTNETS 2010)


MUSIC (subspace based approach)

Experimental Setup

18


Comparison Setup


Each
testing “run” consists of 10 second channel traces.


Random Subset of
4
different radios are selected in each “run” (WiFi,
Bluetooth, ZigBee,
Microwave
) with varying PHY parameters


30 Different “runs” for each signal combination

Evaluation: Classification

0
0.2
0.4
0.6
0.8
1
1.2
-5
-4
-3
-2
-1
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
Single Signal
Classification

DOF

RFDump

Accuracy

SNR

19

DOF achieves greater than 85% accuracy when
the SNR of the detected signal is as low as 0dB

Evaluation: Classification

0
0.2
0.4
0.6
0.8
1
1.2
0
0.01
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
0.2
0.22
0.24
0.26
0.28
0.3
Probability of Missed Classification

Cumulative Fraction

1 Signal

2 Signals

3 Signals

DOF:
Multiple Signal Classification

20

DOF classifies all component signals with greater
than
80%
accuracy, even with 3 interfering signals

Evaluation: Spectrum Occupancy

0
0.1
0.2
0.3
0.4
0.5
0.6
-2
-1
0
1
2
3
4
5
6
7
8
9
10
11
12
13
Single Signal Spectrum
Occupancy
Estimation

Normalized Error

SNR

DOF

Jello

21

DOF’s spectrum occupancy estimates are at least
85% accurate at SNRs as low as 0dB

Evaluation: Spectrum Occupancy

0
0.2
0.4
0.6
0.8
1
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
0.55
0.6
0.65
0.7
0.75
Cumulative Fraction

DOF

Jello

Normalized Error

22

Multiple Signal Spectrum Occupancy Estimation

DOF’s spectrum occupancy estimates are robust
in the presence of multiple overlapping signals

Evaluation: Angle of Arrival

DOF

SecureAngle

(MUSIC)

0.0
0.2
0.4
0.6
0.8
1.0
1.2
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
Cumulative Fraction

Absolute Difference per Angle (Deg)

Multiple Signals: AoA Detection Accuracy

23

In addition to being accurate, DOF can also
associates each AoA with each signal type

Smart Radio Prototype: DOF
-

SR

DOF
-
SR (Policy Aware Smart Radio)


Policy 0


Only use unoccupied spectrum

WiFi

Microwave

Smart Tx

AoA

Freq

2.3
GHz

2.5
GHz

Frequency

2.5
GHz

Smart Rx

AoA

Freq

2.3
GHz

PSD


Policy 1


Use all unoccupied spectrum. Further use spectrum
occupied by microwave ovens.


Policy 2


Use all unoccupied spectrum + microwave occupied
spectrum. Further compete for spectrum occupied by WiFi radios
and get half the time share on that spectrum.


24

Heart Monitor

(ZigBee Based)

DOF
-
SR Performance

0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.5
1.0
Normalized Throughput

Normalized Harm

DOF
-
SR
Policy 0

and
Jello

0.00
0.20
0.40
0.60
0.80
1.00
0.00
0.50
1.00
Normalized Throughput

Normalized Harm

DOF
-
SR
Policy 1

and
Jello

0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.5
1.0
Normalized Throughput

Normalized Harm

DOF
-

SR Policy 2
and Jello

DOF
-
SR

Jello

Legend

DOF
-
SR enables users to decide how
aggressive their policy should be

25

Conclusion

26

DOF exploits repeating patterns to infer 1) type,
2) spectral occupancy, and 3) spatial directions

DOF Performance
Summary


DOF is
robust to SNR of detected
signals


Accurate at received signals as low as 0dB



DOF is
robust to
multiple overlapping signals


Accurate even when three unknown signals are present



DOF is relatively computationally inexpensive


Requires 30% more computation over standard FFT