DOF: A Local Wireless Information Plane Sachin Katti Steven Hong - - PowerPoint PPT Presentation

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 ° 180° AoA Freq 2.3 GHz 2.5 GHz ° 180° Freq 2.3 GHz 2.5 GHz Freq 2.3 GHz 2.5 GHz

4

DOF (Degrees Of Freedom)

Local wireless information plane which provides all 3

• f 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

*𝑼𝒛𝒒𝒇, 𝑮 (𝒋) +𝒐=𝟐

𝑶

DOF Estimation (AoA Detection) ADC

Signal Time Samples

Classification

F( i )

MAC

DOF

*𝑼𝒛𝒒𝒇, 𝑮𝒅, 𝑪𝑿+𝒐=𝟐

𝑶

*𝚰+𝒐=𝟐

𝑶

*𝚰, 𝑼𝒛𝒒𝒇, 𝑮𝒅, 𝑪𝑿+𝒐=𝟐

𝑶

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. 𝑆𝑦

𝛽 𝜐 = 𝑦 𝑜 𝑦∗ 𝑜 − 𝜐 𝑓−𝑘2𝜌𝛽𝑜 ∞ 𝑜

Pattern Frequency (𝛽) – 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

𝑇𝑦

𝛽 𝑔 = 𝑆𝑦 𝛽 𝜐 𝑓−𝑘2𝜌𝑔𝜐 ∞ 𝜐=−∞

= 1 𝑀 𝑌𝑚𝑂 𝑔 𝑌𝑚𝑂

∗ (𝑔 − 𝛽) 𝑀−1 𝑚=0

Pattern Frequency (α) Frequency (f) WiFi Spectral Correlation Function

Feature Vectors are calculated by computing 𝑻𝒚

𝜷 𝒈 at different values of 𝜷

Classifying Signal Type

10

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

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

Feature Dimension 1: 𝑻𝒚

𝜷𝟐 𝒈

Feature Dimension 2: 𝑻𝒚

𝜷𝟑 𝒈

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

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: 𝑻𝒚

𝜷𝟐 𝒈

Feature Dimension 2: 𝑻𝒚

𝜷𝟑 𝒈

Feature Dimension 1: 𝑻𝒚

𝜷𝟐 𝒈

Feature Dimension 2: 𝑻𝒚

𝜷𝟑 𝒈 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 𝑮 𝒋 is more robust than differences in energy

DOF: High Level Architecture

Feature Extraction

*𝑼𝒛𝒒𝒇, 𝑮 (𝒋) +𝒐=𝟐

𝑶

DOF Estimation (AoA Detection) ADC

Signal 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

*𝑼𝒛𝒒𝒇, 𝑮𝒅, 𝑪𝑿+𝒐=𝟐

𝑶

*𝚰+𝒐=𝟐

𝑶

*𝚰, 𝑼𝒛𝒒𝒇, 𝑮𝒅, 𝑪𝑿+𝒐=𝟐

𝑶

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

𝑡 𝑢 = 𝑐𝑑𝑝𝑡 2𝜌𝑔

𝑐𝑢 𝑓𝑘2𝜌𝑔

𝑑𝑢

• 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 𝑔

𝑐 and carrier frequency 𝑔 𝑑 of the signal

1 Bit Sequence b Amplitude modulated Pulse 𝑐𝑑𝑝𝑡 2𝜌𝑔

𝑐𝑢

Pulse multiplied by Carrier Wave 𝑐𝑑𝑝𝑡 2𝜌𝑔

𝑐𝑢 𝑓𝑘2𝜌𝑔

𝑑𝑢

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 𝜷′𝒕 between ,𝒈𝒅 − 𝑪𝑿

𝟑 , 𝒈𝒅 + 𝑪𝑿 𝟑 -

𝒈𝒅, 𝒈𝒅 − 𝑪𝑿

𝟑 , 𝒈𝒅 + 𝑪𝑿 𝟑

𝟑𝒈𝒅 + 𝑪𝑿, 𝟑𝒈𝒅 + 𝑪𝑿

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 𝐺 (𝑗) which can be computed

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

Implementation

RX2 RX 1 RX 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 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

0.2 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

0.2 0.4 0.6 0.8 1 1.2 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

0.1 0.2 0.3 0.4 0.5 0.6

• 2 -1

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

0.2 0.4 0.6 0.8 1 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
• ccupied 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