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 {π―, π»πππ, ππ , π©πΎ}π΅ π=π DOF operates on windows of raw time samples from the ADC {π―}π΅ π=π DOF Estimation (AoA Detection) {π»πππ, ππ , π©πΎ}π΅ π=π DOF Estimation (Spectrum Occupancy) {π»πππ, π (π) }π΅ π=π 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) π π₯πΌ π = π₯ π π₯ ∗ π − π π −π2ππΌπ π Pattern Frequency (πΌ) – 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) ∞ ππ₯πΌ π = π π₯πΌ π π −π2πππ = π=−∞ 1 πΏ πΏ−1 ∗ πππ π πππ (π − πΌ) π=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) πΆ computing πΊπ π at different values of πΆ 9 Classifying Signal Type Feature Dimension 2: πΆ πΊπ π π • Single signals are well separated in the feature vector space, πΉ πΆ Feature Dimension 1: πΊπ π π • Support Vector Machines (SVM) can be used to classify signal type, π 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 πΆ πΊπ π π Feature Dimension 2: Feature Dimension 2: πΆ πΊπ π π • Multiple signals are made up of components and features of single signals, making them difficult to distinguish πΆ πΆ Feature Dimension 1: πΊπ π π Feature Dimension 1: πΊπ π π 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 π π is more robust than differences in energy 12 DOF: High Level Architecture MAC DOF {π―, π»πππ, ππ , π©πΎ}π΅ π=π {π―}π΅ π=π DOF Estimation (AoA Detection) {π»πππ, ππ , π©πΎ}π΅ π=π DOF Estimation (Spectrum Occupancy) The signal types can be leveraged along with the feature vectors to estimate 1) Spectrum Occupancy 2) Spatial Directions {π»πππ, π (π) }π΅ π=π 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 π π‘ = ππππ 2πππ π‘ π π2πππ π‘ 1 Bit Sequence b 0 Amplitude modulated Pulse ππππ 2πππ π‘ Pulse multiplied by Carrier Wave ππππ 2πππ π‘ π π2πππ π‘ • 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 ππ and carrier frequency ππ 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 π©πΎ , ππ π π©πΎ ππ + π WiFi all πΆ′ π between [ππ − + π©πΎ ] π ππ , ππ − , DOFBluetooth leverages this relationship toπ©πΎπcompute the ZigBee πππ + π©πΎ, πππ +type π©πΎ 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 πΉ(π) 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