DOF:
A Local Wireless Information Plane
Steven Hong
Sachin Katti
Stanford University
August 17, 2011
1
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?
Knowledge of
1. The protocol types operating in the local vicinity
2. The spectrum occupancy of each type
Heart Monitor
0
°
0
°
AoA
AoA
180°
WiFi AP
180°
3. The spatial directions of each type
2.3
2.3
GHz
GHz
Freq
2.5
GHz
2.3
2.3
GHz
GHz
Smart Transmitter
Smart Receiver
Microwave
4
Freq
2.5
2.5
GHz
GHz
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
MAC
DOF
{𝚯, 𝑻𝒚𝒑𝒆, 𝑭𝒄 , 𝑩𝑾}𝑵
𝒏=𝟏
DOF operates on windows of raw
time samples from the ADC
{𝚯}𝑵
𝒏=𝟏
DOF Estimation
(AoA Detection)
{𝑻𝒚𝒑𝒆, 𝑭𝒄 , 𝑩𝑾}𝑵
𝒏=𝟏
DOF Estimation
(Spectrum Occupancy)
{𝑻𝒚𝒑𝒆, 𝑭 (𝒊) }𝑵
𝒏=𝟏
Classification
F( i )
Feature
Extraction
Time Samples
Signal
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
6
Key Insight
For almost all “man-made” signals – there are hidden repeating
patterns that are unique and necessary for operation
CP
Data
CP
CP
Data
Data
…………………….
Repeating Patterns in WiFi OFDM signals
Time
Repeating Patterns in Zigbee signals
Leverage unique patterns to infer 1) type, 2)
spectral occupancy, and 3) spatial directions
7
Extracting Features from Patterns
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.
∞
Cyclic Autocorrelation Function (CAF) 𝑅𝑥𝛼 𝜏 =
𝑥 𝑛 𝑥 ∗ 𝑛 − 𝜏 𝑒 −𝑗2𝜋𝛼𝑛
𝑛
Pattern Frequency (𝛼) – The frequency at which the pattern repeats
Advantages
•Robustness to noise,
•Uniqueness for each
protocol
Disadvantage: Computationally expensive to
calculate
the
patterns
in
this
manner
Delay (τ)
Pattern Frequency (α)
8
Feature Extraction:
Efficient Computation
The CAF can be represented using an equivalent form called the Spectral
Correlation Function (SCF)
∞
𝑆𝑥𝛼 𝑓 =
𝑅𝑥𝛼 𝜏 𝑒 −𝑗2𝜋𝑓𝜏
=
𝜏=−∞
1
𝐿
𝐿−1
∗
𝑋𝑙𝑁 𝑓 𝑋𝑙𝑁
(𝑓 − 𝛼)
𝑙=0
WiFi Spectral Correlation Function
• SCF can be calculated
for Discrete Time
Windows using just
FFTs
Feature Vectors are calculated by
Frequency (f)
Pattern Frequency (α)
𝜶
computing 𝑺𝒙 𝒇 at different values of 𝜶
9
Classifying Signal Type
Feature Dimension 2:
𝜶
𝑺𝒙 𝟐 𝒇
• Single signals are well separated in the feature vector space, 𝐹
𝜶
Feature Dimension 1: 𝑺𝒙 𝟏 𝒇
•
Support Vector Machines (SVM) can be used to classify signal type, 𝑇
Works well when there is a single signal but fails
when there are multiple interfering signals
10
Multiple interfering signals are not
straightforward to classify
𝜶
𝑺𝒙 𝟐 𝒇
Feature Dimension 2:
Feature Dimension 2:
𝜶
𝑺𝒙 𝟐 𝒇
• Multiple signals are made up of components and features of single
signals, making them difficult to distinguish
𝜶
𝜶
Feature Dimension 1: 𝑺𝒙 𝟏 𝒇
Feature Dimension 1: 𝑺𝒙 𝟏 𝒇
Need a robust algorithm to determine the
number of interfering signals
11
Inferring the number of signals:
Exploiting Asynchrony
1) Real signal packets are asynchronous
2) This asynchrony shows up in F( i ) as an increase or decrease in the
number of non-zero components
Received Signal
ZigBee
Overlapping Packets
WiFi
Nonzero Components in
F( i )
t
Measuring differences in 𝑭 𝒊 is more robust
than differences in energy
12
DOF: High Level Architecture
MAC
DOF
{𝚯, 𝑻𝒚𝒑𝒆, 𝑭𝒄 , 𝑩𝑾}𝑵
𝒏=𝟏
{𝚯}𝑵
𝒏=𝟏
DOF Estimation
(AoA Detection)
{𝑻𝒚𝒑𝒆, 𝑭𝒄 , 𝑩𝑾}𝑵
𝒏=𝟏
DOF Estimation
(Spectrum Occupancy)
The signal types can be leveraged along with
the feature vectors to estimate
1) Spectrum Occupancy
2)
Spatial Directions
{𝑻𝒚𝒑𝒆, 𝑭 (𝒊) }𝑵
𝒏=𝟏
Feature
Extraction
DOF = Active
Asynchrony
Detector/
Power
Normalization
If ΔL0<-Threshold
Time Samples
Signal
Counter++
Counter-N Signals
ADC
13
SVM-1
Sig1 Class
Sig i Class
SVM-N
SigN Class
...
If ΔL0>Threshold
...
F( i )
While
...
Classification
1 Signal
Estimating Spectrum Occupancy
• Communication signals are sequences of periodic pulses
𝑠 𝑡 = 𝑏𝑐𝑜𝑠 2𝜋𝑓𝑏 𝑡 𝑒 𝑗2𝜋𝑓𝑐 𝑡
1
Bit Sequence
b
0
Amplitude modulated Pulse
𝑏𝑐𝑜𝑠 2𝜋𝑓𝑏 𝑡
Pulse multiplied by Carrier Wave
𝑏𝑐𝑜𝑠 2𝜋𝑓𝑏 𝑡 𝑒 𝑗2𝜋𝑓𝑐 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 𝑓𝑏 and carrier frequency 𝑓𝑐 of the signal
14
Estimating Spectrum Occupancy
• Because these patterns repeat, they are natural components of the
ZigBee Spectral Correlation Function
feature vector
Modulated Zigbee Signal
Time
Frequency (f)
Pattern Frequency (α)
Relationship between feature vector and Bandwidth/Carrier Frequencies
Signal Type
Feature Vector Frequencies
𝑩𝑾
, 𝒇𝒄
𝟐
𝑩𝑾
𝒇𝒄 + 𝟐
WiFi
all 𝜶′ 𝒔 between [𝒇𝒄 −
+
𝑩𝑾
]
𝟐
𝒇𝒄 , 𝒇𝒄 −
,
DOFBluetooth
leverages this relationship
to𝑩𝑾𝟐compute
the
ZigBee
𝟐𝒇𝒄 + 𝑩𝑾,
𝟐𝒇𝒄 +type
𝑩𝑾
spectral occupancy of each
signal
15
Estimating Angles of Arrival
Incoming Signal
...
...
Array Elements
1
d
2
M
• Each array element experiences a delay of τ relative to the first
array element, which is a function of the Angle of Arrival (AoA)
DOF uses the same feature vector to infer
• This unique delay induces a particular characteristic on the
1)type,
2)spectral occupancy, 3)spatial directions
feature vector 𝐹(𝑖) which can be computed
16
Implementation
RX 1
RX 3
RX2
• 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
Experimental Setup
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
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)
18
Evaluation: Classification
1.2
Single Signal Classification
Accuracy
1
0.8
0.6
0.4
DOF
RFDump
0.2
0
-5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
SNR
DOF achieves greater than 85% accuracy when
the SNR of the detected signal is as low as 0dB
19
1.2
DOF: Multiple Signal Classification
1
0.8
0.6
1 Signal
2 Signals
3 Signals
0.4
0.2
0
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
Cumulative Fraction
Evaluation: Classification
Probability of Missed Classification
DOF classifies all component signals with greater
than 80% accuracy, even with 3 interfering signals
20
Evaluation: Spectrum Occupancy
Normalized Error
0.6
Single Signal Spectrum Occupancy Estimation
0.5
0.4
0.3
DOF
Jello
0.2
0.1
0
-2 -1 0 1 2 3 4 5 6 7 8 9 10 11 12 13
SNR
DOF’s spectrum occupancy estimates are at least
85% accurate at SNRs as low as 0dB
21
Multiple Signal Spectrum Occupancy Estimation
1
0.8
0.6
0.4
DOF
Jello
0.2
0.75
0.7
0.65
0.6
0.55
0.5
0.45
0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
0
Cumulative Fraction
Evaluation: Spectrum Occupancy
Normalized Error
DOF’s spectrum occupancy estimates are robust
in the presence of multiple overlapping signals
22
Evaluation: Angle of Arrival
Multiple Signals: AoA Detection Accuracy
Cumulative Fraction
1.2
1.0
0.8
0.6
DOF
SecureAngle
(MUSIC)
0.4
0.2
0.0
4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
Absolute Difference per Angle (Deg)
In addition to being accurate, DOF can also
associates each AoA with each signal type
23
Smart Radio Prototype: DOF- SR
DOF-SR (Policy Aware Smart Radio)
• Policy 0 – Only use unoccupied spectrum
• 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.
WiFi
AoA
PSD
Frequency
2.3
GHz
Freq
2.5
GHz
Microwave
24
Smart Rx
AoA
Smart Tx
Heart Monitor
(ZigBee Based)
2.3
GHz
Freq
2.5
GHz
Legend
DOF-SR Performance
0.8
0.6
0.4
0.2
0.0
0.0
0.5
1.0
Normalized Harm
1.00
0.80
0.60
0.40
0.20
0.00
0.00
0.50
1.00
Normalized Harm
DOF - SR Policy 2
and Jello
Normalized Throughput
1.0
DOF-SR Policy 1
and Jello
Normalized Throughput
Normalized Throughput
DOF-SR Policy 0
and Jello
DOF-SR
Jello
1.0
0.8
0.6
0.4
0.2
0.0
0.0
0.5
1.0
Normalized Harm
DOF-SR enables users to decide how
aggressive their policy should be
25
Conclusion
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
26