CSIsnoop: Attacker Inference of Channel State Information in Multi-User WLANs

Transcription

1 CSIsnoop: Attacker Inference of Channel State Information in Multi-User WLANs Xu Zhang and Edward W. Knightly ECE Department, Rice University

2 Channel State Information (CSI) CSI plays a key role in wireless networks 2

3 Channel State Information (CSI) CSI plays a key role in wireless networks H 1 H 1 = A 1 e jθ 1 A 2 e jθ 2 A 3 e jθ 3 > λ/2 uncorrelated H 2 H 2 = B 1 e jφ 1 B 2 e jφ 2 B 3 e jφ 3 2

4 Channel State Information (CSI) CSI plays a key role in wireless networks Increase throughput e.g., IEEE ac H 1 H 1 = A 1 e jθ 1 A 2 e jθ 2 A 3 e jθ 3 > λ/2 uncorrelated H 2 H 2 = B 1 e jφ 1 B 2 e jφ 2 B 3 e jφ 3 3

5 Channel State Information (CSI) CSI plays a key role in wireless networks Increase throughput e.g., IEEE ac Enhance security e.g., CSI-based password H 1 Meansure H 1 Quantify Password bits Quantify Meansure H 1 Password bits 4

6 Channel State Information (CSI) Conventionally, any nodes half-a-wavelength away cannot guess the laptop s CSI H 1 Meansure H 1 Quantify Password bits > λ/2 Quantify Meansure H 1 Password bits What is H 1? Adversary 4

7 Channel State Information (CSI) Conventionally, any nodes half-a-wavelength away cannot guess the laptop s CSI However, we show that even a passive adversary can actually infer the laptop s CSI H 1 Meansure H 1 Quantify Password bits > λ/2 Meansure H 1 H 2 Quantify Password bits Infer H 1! Adversary 4

8 CSIsnoop A fundamental conflict between using CSI to optimize PHY and hiding CSI from adversaries Secure CSI Increase throughput 5

9 Roadmap Threat Model CSIsnoop Framework Implementation on WARP and Experimental Evaluation 6

10 Threat Model Legitimate Clients A typical multi-user WLAN with OFDM transmission Multi-antenna AP Single-antenna clients Bob 1 to Bob n always uses all her antennas to boost the throughput Bob 1... Bob n 7

11 Threat Model Legitimate Clients A typical multi-user WLAN with OFDM transmission Explicit channel sounding like IEEE ac Bob 1... Bob n 8

12 Threat Model Legitimate Clients A typical multi-user WLAN with OFDM transmission Explicit channel sounding like IEEE ac Encrypted feedback or sounding sequence Bob 1... Bob n (1) Encrypted feedback (2) Encrypted sounding sequence Sounding sequence Encrypted feedback Encrypted sounding sequence Feedback Bob 1 Bob 1 8

13 Threat Model Legitimate Clients A typical multi-user WLAN with OFDM transmission Explicit channel sounding like IEEE ac Encrypted feedback or sounding sequence Zero-force beamforming, but CSIsnoop can be generalized to other beamforming algorithms Bob 1... Bob n (1) Encrypted feedback (2) Encrypted sounding sequence Sounding sequence Encrypted feedback Encrypted sounding sequence Feedback Bob 1 Bob 1 8

14 Threat Model Adversary Eve is a multi-antenna adversary Same number of antennas as the AP Bob 1... Bob n Eve 9

15 Threat Model Adversary Eve is a multi-antenna adversary Same number of antennas as the AP In range of Knows which Bobs are included in multi-user beamforming transmission Knows part of the symbols in each Bob s downlink data packets Bob 1... Bob n Eve 9

16 Threat Model Adversary Eve is a multi-antenna adversary Same number of antennas as the AP In range of Knows which Bobs are included in multi-user beamforming transmission Knows part of the symbols in each Bob s downlink data packets Packet from to Bob: Bob 1... Bob n Header Payload Eve Eve knows these symbols before overhearing them 9

17 CSIsnoop Framework Overhear CSIsnoop Known symbols at Eve Eve s transmit beamforming weights Bob 1 Bob 3 Bob 2 CSI between and each Bob 10

18 CSIsnoop Framework W A, P, X H AB1 H ABm H AE Y Bob 1 Eve Bob m... Bob n Y = H AE W A P X + N 11

19 CSIsnoop Framework W A, P, X H AB1 H ABm H AE Y Bob 1 Eve Bob m... Bob n Data to selected Bobs Signals at Eve Y = H AE W A P X + N Diagonal transmit power scaling matrix Transmit beamforming weights 11

20 CSIsnoop Framework Y = H AE W A P X + N (1) H AB W A = I (2) 12

21 CSIsnoop Framework Y = H AE W A P X + N (1) H AB W A = I (2) Known-transmitted-symbol attack X: known symbols at Eve W E = arg min Y W E X Y: overheard signals at Eve 12

22 CSIsnoop Framework Y = H AE W A P X + N (1) H AB W A = I (2) Known-transmitted-symbol attack X: known symbols at Eve W E = arg min Y W E X H AE W A P W E Y: overheard signals at Eve 12

23 CSIsnoop Framework Y = H AE W A P X + N (1) H AB W A = I (2) Known-transmitted-symbol attack X: known symbols at Eve First assume that Eve knows H AE W E = arg min Y W E X H AE W A P W E W A P H 1 AE W E Y: overheard signals at Eve 12

24 CSIsnoop Framework Y = H AE W A P X + N (1) H AB W A = I (2) Known-transmitted-symbol attack Eve computes W A P Eve does not know P and cannot solve W A 13

25 CSIsnoop Framework Y = H AE W A P X + N (1) H AB W A = I (2) Known-transmitted-symbol attack Eve computes W A P Eve does not know P and cannot solve W A and Bob use span(h ABj ) instead of H ABj Remove inter-user interference transmits signals of Bob i j into null(h ABj ) CSI-based password Normalize H ABj as and Bob j may use different transmit power 13

26 CSIsnoop Framework Y = H AE W A P X + N (1) H AB W A = I (2) H AB1 H AB2 H ABm W A1 W A2 W Am = I To compute span(h ABj ), Eve only needs to know the direction of each column of W A W A P preserves the direction of each column of W A W A P = W A1,, W Am p 1 p m = [W A1 p 1,, W Am p m ] 14

27 CSIsnoop Framework Y = H AE W A P X + N (1) H AB W A = I (2) Known-transmitted-symbol attack H AB1 H AB2 H ABm W A1 W A2 W Am = I Estimate span(h ABj ) If the number of selected Bobs = s antenna number span(h ABj ) can be determined If the number of selected Bobs < s antenna number Eve overhears > 1 beamforming transmissions to compute span(h ABj ) 15

28 Eve Estimates Her Channel H AE Eve needs to estimate her channel H AE 16

29 Eve Estimates Her Channel H AE Eve needs to estimate her channel H AE (1) Encrypted feedback (2) Encrypted sounding sequence Sounding sequence Encrypted feedback Encrypted sounding sequence Feedback Bob 1 Bob 1 16

30 Eve Estimates Her Channel H AE Eve needs to estimate her channel H AE (1) Encrypted feedback (2) Encrypted sounding sequence Sounding sequence Encrypted feedback Encrypted sounding sequence Feedback Bob 1 Bob 1 Eve estimates H AE through the unencrypted channel sounding sequence 16

31 Eve Estimates Her Channel H AE Eve needs to estimate her channel H AE (1) Encrypted feedback (2) Encrypted sounding sequence Sounding sequence Encrypted feedback Encrypted sounding sequence Feedback Bob 1 Bob 1 Eve estimates H AE through the unencrypted channel sounding sequence Eve can still estimate H AE by using the dynamic cyclic shift 16

32 Implementation on WARP A multi-user MIMO WLAN in the 5 GHz band A 4-antenna WARP as 4 single-antenna WARPs as the Bobs ac packet format ac-like multi-user beamforming Bob 1 Bob 2 Bob 3 Bob 4 17

33 Implementation on WARP A multi-user MIMO WLAN in the 5 GHz band A 4-antenna WARP as 4 single-antenna WARPs as the Bobs ac packet format ac-like multi-user beamforming Encrypted channel sounding [CSIsec, CCS 2014] Bob 2 Bob 1 Bob 3 Bob 4 17

34 Implementation on WARP A multi-user MIMO WLAN in the 5 GHz band A 4-antenna WARP as 4 single-antenna WARPs as the Bobs ac packet format ac-like multi-user beamforming Encrypted channel sounding [CSIsec, CCS 2014] CSIsnoop at Eve Same number of antennas as Correct timing offset/carrier frequency offset Estimate H AE Use CSIsnoop to compute Bobs CSI Bob 2 Bob 1 H AE Bob 3 Bob 4 Eve 17

35 Experimental Evaluation Setup Configure and Eve to have 2, 3, or 4 antennas Collect >100,000 rounds of over-the-air transmissions in different indoor environments 18

36 Experimental Evaluation Setup Configure and Eve to have 2, 3, or 4 antennas Collect >100,000 rounds of over-the-air transmissions in different indoor environments Bob 1 Bob 2 Bob 3 Bob 4 18

37 Experimental Evaluation Setup Configure and Eve to have 2, 3, or 4 antennas Collect >100,000 rounds of over-the-air transmissions in different indoor environments Eve Bob 1 Bob 2 Bob 3 Bob 4 CSIsnoop 18

38 Experimental Evaluation Setup Configure and Eve to have 2, 3, or 4 antennas Collect >100,000 rounds of over-the-air transmissions in different indoor environments Eve Bob 1 Bob 2 Bob 3 Bob 4 CSIsnoop Metric Normalized correlation c between Bob s measured CSI and Eve s computed CSI c = 1 indicates that the measured CSI and the computed CSI are perfectly correlated 18

39 Estimation Accuracy of CSIsnoop 1 CCDF Eve does not use CSIsnoop Eve cannot estimate Bob s CSI by directly using her own CSI Correlation c 19

40 Estimation Accuracy of CSIsnoop CCDF 0.8 CCDF / 3/ 4 ants,h AE ants,h AB ;comp 3 ants,h AB ;comp 4 ants,h AB ;comp Correlation c ants,h AB ;comp 3 ants,h AB ;comp 4 ants,h AB ;comp Correlation c Average correlation c increases from 0.46 to over 0.99 with CSIsnoop Estimation accuracy reduces when has more antennas 19

41 Impact of Eve s Channel H AE Estimation accuracy is closely related to Eve s SNR and cond(h AE ) cond(h AE ) is the ratio between the largest and smallest singular value of H AE In the previous slide, average SNR is 30 db and cond H AE = 5 20

42 Impact of Eve s Channel H AE Estimation accuracy is closely related to Eve s SNR and cond(h AE ) cond(h AE ) is the ratio between the largest and smallest singular value of H AE In the previous slide, average SNR is 30 db and cond H AE = 5 Reducing SNR 1 Correlation c Signal SNR at Eve (db) 20

43 Impact of Eve s Channel H AE Estimation accuracy is closely related to Eve s SNR and cond(h AE ) cond(h AE ) is the ratio between the largest and smallest singular value of H AE In the previous slide, average SNR is 30 db and cond H AE = 5 Reducing SNR Increasing cond(h AE ) 1 1 Correlation c Correlation c Signal SNR at Eve (db) Average cond(h AE ) 20

44 More Overheard Packets When Eve s SNR is small and cond H AE is large Eve can overhear more packets to increase her estimation accuracy SimpAvg Compute the average of the several computed CSI SubSpaceSearch Compute the most likely 1-dimensional sub-space spanned by the several computed CSI 21

45 More Overheard Packets Correlation c SubSpaceSeach SimpAvg Number of Packets Eve s SNR is 20 db and cond H AE = 30 SubSpaceSearch increases estimation accuracy while SimpAvg may even reduce it 22

46 More Overheard Packets Fractional timing offset due to ADC sampling at Eve A maximum error of T/2 in determining the start of each overheard packet Unknown phase rotation for Eve s computed CSI for each overheard packet SubSpaceSearch will not be influenced by the unknown phase rotation 23

47 CSI-Based Attacks After Eve infers Bob s CSI Eve can compute over 85% of the CSI-based password between and Bob Eve can selectively jam and thus only reduce the uplink throughput of a target Bob 24

48 Summary A fundamental conflict between using CSI to boost throughput and hiding CSI Describe the CSIsnoop framework Experimental results show high estimation accuracy of CSIsnoop A more careful examination of using CSI as a shared secret Design schemes to detect and prevent attacks based on CSI 25