1. Advancing Wireless Link Signatures for Location Distinction Mobicom 2008 Junxing Zhang, Mohammad H. Firooz Neal Patwari, Sneha K. Kasera University of Utah Salt Lake City, USA

2. Outline Introduction Multipath-Based Link Signatures – Multiple Tone Probing – Temporal CIR Signature Innovative Methods – Refined Metric for multiple Tone Signatures – Complex Temporal Signature Framework for Location Distinction Quantitative Comparisons of Link Signatures Temporal Behavior of Link Signatures Conclusion

3. Introduction Location Distinction v.s. Localization Proposed wireless link signatures – RSS – Channel Gains of Multi-tonal Probes – Temporal Channel Impulse Response The RSS-based method has a consistently lower detection rate and a higher false alarm rate. Location DistinctionLocalization Can thrive in multipath channelSuffer from inaccuracies by multipath More sensitive to motion (less than 1 meter) May not be able to determine within a meter of accuracy Less coverageLarger coverage

4. Multipath-Based Link Signatures Two existing multipath-base link signatures and metrics – Multiple tone Probing – Temporal CIR Signature Comparison and discussion

5. Multipath Channel Response Multipath caused by – Reflections, diffractions and scattering of the radio waves. – Time-delayed, attenuated and phase-shifted Impulse response of multi-path fading channel – Time-variant: – Time-invariant: (for a packet duration) Received signal – Path number Phase shift Magnitude gain Channel frequency response Channel impulse response (CIR)

6. Multipath Channel Response Recover CIR from received signal Ps : the power of the sent signal inside the band

7. Multiple Tone Probing Frequency domain Measures frequency response of multiple carriers – K carrier waves are simultaneously transmitted to the receiver The nth recorded multiple tone signature of the link between transmitter i and receiver j is: – f K : the carrier frequency of the Kth carrier wave {H(f K )} used as the multiple tone signature

8. Multiple Tone Probing Metric – The Nth multiple tone signature h (N) is compared with each previously measured signature in the history H i,j using average correlation statistic. Measurement of similarity – Low -> different, high -> similare Correlation of the nth and the Nth measurements Average squared magnitude of the elements of

9. Temporal CIR Signature Time domain Estimation of the impulse responses a function of time delay and magnitude Metric – Minimum normalized Euclidean distance – Difference: Low -> similar The nth sampled link signature measurement of the link between transmitter I and receiver j Tr the sampling interval at the receiver and S+1 is the number of the samples

10. Comparison and Discussion Qualitative comparison – Temporal signature can be more robust against small changes in multipath. – The inclusion of phase information in multiple tone signature effectively increases the richness of the measurement space. – The temporal link signature has the advantage of operation in the time domain which de-correlates multipath at different delays – The multiple tone link signature has the advantage of using a complex- valued signature which preserves phase information

11. Complex Temporal Signature Proposed method – Combines the best features of both the temporal link signature method and the multiple tone probing method. – enhanced signature: – The complex link signature retains phase information in a manner similar to the multiple tone link signature. Comparing to temporal link signature, the magnitude of each gain is not taken

12. Issue: Phase Changes Random Phase Shift: – Some phase changes in the link signature have nothing to do with any changes in the link – Clock or carrier frequency shifts – h -> Time offset between clocks Different carrier frequency in receiver and transmitter

13. Issue: Phase Changes Given 2 complex temporal link signatures h and g – Represent the shift-removed difference with a new Φ 2 difference – The Φ 2 difference, which minimizes the random phase shift between two measurements before calculating distance, can be efficiently and explicitly calculated using simple vector operations.

14. Quantitative Comparisons of Link Signature Comparison: – Multiple tone probing – Temporal channel impulse response – Complex temporal link signatures Obtained from CRAWDAD 5 measurement: 4 for history, 1 for test ROC plot: receiver operating characteristic – How the probability of detection varies with the probability of false alarm.

15. Normalized Metric – Normalizing each multiple tone signature to its magnitude before the calculation of the correlation statistic – Each channel frequency response is normalized to the square root of its average power. – Avoiding missed detections, when higher h (N) occurs. =

16. Normalized Metric ROC curves of comparing performance of original and the normalized metrics in the multiple tone probing method

17. Framework for location distinction N-1 link signaturesMetric Location changed Threshold Include it in Discard the oldest Performance Evaluation P FA : the probability of false alarm P D : the probability of detection γ

18. Comparison (1) Multiple Tone v.s. Temporal Link Signatures K increases -> improvement Coherence bandwidth -> not separate enough -> correlation

19. Comparison (2) Three methods

20. Multiple Receiver Performance Multiple tone Complex temporal

21. Temporal Behavior of Link Signatures Temporal changes in link behavior can significantly increase the probability of false alarms. LOCATION A,B,C,D Record response vectors comprising 600 complex temporal responses. Each impulse response is a vector of 100 complex numbers

22. Observation Isomap 2D embedding coordinates – Non-linear dimensionality reduction to reduce the 100 dimension vectors to just 1-2 dimension

23. Markov Model Use 1-D embedding of the Isomap algorithm – Like an amplitude modulation signal -> use AM demodulator to capture the envelope of the pattern

24. False alarm case Different-State False Alarm (DSFA) – A link signature is measured in state I, but no signature previously measured in state I exists in the history. False alarm is raised. Policies of buffer replacement – Policy 1: The history has a FIFO replacement policy. – Policy 2: The history is subdivided into K separate FIFO buffers. One for each state in the Markov chain

25. Two-state Markov Chain Model Policy 1: evaluate the probability of DFSA – (1) given we are entering state 1, the probability that we stayed in state 2 N time units ago – (2) given we are entering state2, the probability that we stayed in state 1 N time units ago

26. Results The rate of convergence is very slow By using Policy 2, we see virtually no DSFA errors.

27. Conclusion Compare two existing multipath-based location distinction methods Improve the multiple tone probing method Develop a new link signature outperforms the existing two A measurement campaign to understand and model the temporal behavior of link signatures to reduce the probaility of false alarms