DOF: A Local Wireless Information Plane

Stanford University

Steven Hong Sachin Katti

1

August 17, 2011

ProblemUnlicensed 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 type3. The spatial directions of each type

Knowledge of

WiFi AP Heart Monitor

AoA

Freq2.3 GHz

2.5 GHz

0 °18

0°

AoA

Freq2.3 GHz

2.5 GHz

0 °18

0°

Freq2.3 GHz

2.5 GHzFreq

2.3 GHz

2.5 GHz

4

DOF(Degrees Of Freedom)

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

FeatureExtraction

{𝑻𝒚𝒑𝒆 , �⃗� (𝒊)}𝒏=𝟏𝑵

DOF Estimation(AoA Detection)

ADCSignal

Time Samples

Classification

F( i )⃗

MAC

DOF

{𝑻𝒚𝒑𝒆 ,𝑭 𝒄 ,𝑩𝑾 }𝒏=𝟏𝑵

{𝚯 }𝒏=𝟏𝑵

{𝚯 ,𝑻𝒚𝒑𝒆 ,𝑭𝒄 ,𝑩𝑾 }𝒏=𝟏𝑵

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 detect1. Signal Type2. Spectral Occupancy3. 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 CPData Data Data …………………….

Repeating Patterns in WiFi OFDM signals

Repeating Patterns in Zigbee signalsTime

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.

𝑅𝑥𝛼 (𝜏 )=∑

𝑛

∞

𝑥 [𝑛 ] [𝑥∗ [𝑛−𝜏 ] ]𝑒− 𝑗2 𝜋𝛼𝑛

Pattern Frequency () – 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

𝑆𝑥𝛼 ( 𝑓 )= ∑

𝜏=−∞

∞

𝑅𝑥𝛼 (𝜏 )𝑒− 𝑗 2𝜋 𝑓 𝜏¿ 1

𝐿∑𝑙=0

𝐿−1

𝑋 𝑙𝑁 ( 𝑓 ) 𝑋 𝑙𝑁∗ ( 𝑓 −𝛼)

Pattern Frequency (α)Frequency (f)

WiFi Spectral Correlation Function

Feature Vectors are calculated by computing at different values of

Classifying Signal Type

10

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

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

Feature Dimension 1:

Fea

ture

Dim

ens

ion

2:

• 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:

Fea

ture

Dim

ens

ion

2:

Feature Dimension 1:

Fea

ture

Dim

ens

ion

2:

11

1) Real signal packets are asynchronous

12

Inferring the number of signals: Exploiting Asynchrony

ZigBee

t

Overlapping PacketsWiFi

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 is more robust than differences in energy

DOF: High Level Architecture

FeatureExtraction

{𝑻𝒚𝒑𝒆 , �⃗� (𝒊)}𝒏=𝟏𝑵

DOF Estimation(AoA Detection)

ADCSignal

Time Samples

Classification

F( i )⃗

MAC

Asynchrony Detector/

Power Normalization

SVM-1

SVM-N

Counter++

. . .

If ΔL0>Threshold

Counter--If ΔL0<-Threshold

Sig1 Class

Sig i Class

1 Signal

N Signals

SigN Class

. . .

. . .

While

DOF = Active

DOF

{𝑻𝒚𝒑𝒆 ,𝑭 𝒄 ,𝑩𝑾 }𝒏=𝟏𝑵

{𝚯 }𝒏=𝟏𝑵

{𝚯 ,𝑻𝒚𝒑𝒆 ,𝑭𝒄 ,𝑩𝑾 }𝒏=𝟏𝑵

The signal types can be leveraged along with the feature vectors to estimate 1) Spectrum Occupancy

2) Spatial DirectionsDOF Estimation

(Spectrum Occupancy)

13

Estimating Spectrum Occupancy

14

• Communication signals are sequences of periodic pulses

𝑠 (𝑡 )=𝑏𝑐𝑜𝑠 (2𝜋 𝑓 𝑏𝑡 )𝑒 𝑗 2𝜋 𝑓 𝑐𝑡

• 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

Pulse multiplied by Carrier Wave

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 between

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

Θd si

nΘ

• This unique delay induces a particular characteristic on the feature vector which can be computed

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

Implementation

RX2

RX 1RX 3

• Channel traces 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 ApproachesIdentifying 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

-5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 150

0.2

0.4

0.6

0.8

1

1.2 Single Signal Classification

DOFRFDump

Acc

ura

cy

SNR

19

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

Evaluation: Classification

00.02

0.06 0.10.14

0.180.22

0.26 0.30

0.2

0.4

0.6

0.8

1

1.2

Probability of Missed Classification

Cu

mu

lati

ve F

ract

ion

1 Signal2 Signals3 Signals

DOF: Multiple Signal Classification

20

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

Evaluation: Spectrum Occupancy

-2 -1 0 1 2 3 4 5 6 7 8 9 10 11 12 130

0.1

0.2

0.3

0.4

0.5

0.6 Single Signal Spectrum Occupancy EstimationN

orm

aliz

ed E

rro

r

SNR

DOFJello

21

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

Evaluation: Spectrum Occupancy

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 0.750

0.2

0.4

0.6

0.8

1

Cu

mu

lati

ve F

ract

ion

DOFJello

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

DOFSecureAngle

(MUSIC)

4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 210.0

0.2

0.4

0.6

0.8

1.0

1.2Multiple Signals: AoA Detection Accuracy

Absolute Difference per Angle (Deg)

Cum

ulati

ve F

racti

on

23

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

Smart Radio Prototype: DOF- SRDOF-SR (Policy Aware Smart Radio)

• Policy 0 – Only use unoccupied spectrum

WiFi

Microwave

Smart Tx

Ao

A

Freq2.3 GHz

2.5 GHz

Frequency

2.5 GHz

Smart Rx

Ao

AFreq

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 1.00.0

0.2

0.4

0.6

0.8

1.0

DOF-SR Policy 0 and Jello

Normalized Harm

Nor

mal

ized

Thr

ough

put

0.00 0.50 1.000.00

0.20

0.40

0.60

0.80

1.00

DOF-SR Policy 1 and Jello

Normalized Harm

Nor

mal

ized

Thr

ough

put

0.00.20.40.60.81.00.0

0.2

0.4

0.6

0.8

1.0

DOF - SR Policy 2 and Jello

Normalized Harm

Nor

mal

ized

Thr

ough

put

DOF-SRJello

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