JICE: Joint Data Compression and Encryption for Wireless Energy Auditing Networks

Transcription

1 JICE: Joint Data Compression and Encryption for Wireless Energy Auditing Networks Sheng-Yuan Chiu 1,2, Hoang Hai Nguyen 1, Rui Tan 1, David K.Y. Yau 1,3,Deokwoo Jung 1 1 Advanced Digital Science Center, Illinois at Singapore 2 National Tsing Hua University, Taiwan 3 Singapore University of Technology and Design, Singapore

2 Outline Motivation Design of JICE Secrecy of JICE Eperiment 2/19

3 Wireless Energy Auditing Buildings account for 40% electricity use Wireless appliance submetering Smart plugs (ZigBee radio) 3/19

4 Wireless Energy Auditing Buildings account for 40% electricity use Wireless appliance submetering Efficiency analysis 56% energy wasted in our office[jung 2013] Smart plugs (ZigBee radio) 3/19

5 Objectives & Challenges 50W 60W 30W 25W 10W 50W 70W 58W 50W 55W 50W ZigBee smart plugs 90W 50W ZigBee base station 4/19

6 Objectives & Challenges 50W 60W 30W 25W 10W 50W 70W 58W 50W 55W 50W ZigBee smart plugs 90W 50W ZigBee base station Increase coverage (# of meters) and sampling rate 10% coverage by 455 plugs [Haggerty 2012] Down to 1Hz to support load profiling 4/19

7 Objectives & Challenges Crowded wireless band Heavy WiFi traffic 50W 60W 30W 25W 10W 50W 70W 58W 50W 55W 50W ZigBee smart plugs 90W 50W ZigBee base station Increase coverage (# of meters) and sampling rate 10% coverage by 455 plugs [Haggerty 2012] Down to 1Hz to support load profiling 4/19

8 Objectives & Challenges Crowded wireless band Heavy WiFi traffic 50W 60W 30W 25W 10W 50W 70W 58W 50W 55W 50W ZigBee smart plugs 90W 50W ZigBee base station Increase coverage (# of meters) and sampling rate 10% coverage by 455 plugs [Haggerty 2012] Down to 1Hz to support load profiling Data secrecy during wireless communication Threat model: wireless eavesdropping 4/19

9 Objectives & Challenges Costly encryption No crypto for smart meters [Rouf 2012] Crowded wireless band Heavy WiFi traffic 50W 60W 30W 25W 10W 50W 70W 58W 50W 55W 50W ZigBee smart plugs 90W 50W ZigBee base station Increase coverage (# of meters) and sampling rate 10% coverage by 455 plugs [Haggerty 2012] Down to 1Hz to support load profiling Data secrecy during wireless communication Threat model: wireless eavesdropping 4/19

10 Conventional Scheme (Pipeline) Radio sensor compress encrypt decrypt decompress Power meter Base station reduce bandwidth use prevent eavesdropping Inefficient for resource-constrained plugs Computation-intensive compressor and cipher 5/19

11 Compressive Sensing Random matri Smart plug = Compressed Original Efficient compression Simple matri multiplication Most computation to recovery side Weakly encrypt signal [Rachlin 2008] Shared secret random matri Radio Base station Recovery Constrained optimization 6/19

12 Compressive Sensing Random matri Smart plug O(n) = Compressed Original Efficient compression Simple matri multiplication Most computation to recovery side Weakly encrypt signal [Rachlin 2008] Shared secret random matri Radio Base station Recovery Constrained optimization 6/19

13 Outline Motivation Design of JICE Secrecy of JICE Eperiment 7/19

14 Compressive Sensing Basics = Compression y M 1 M N N 1 M N 8/19

15 Compressive Sensing Basics = Compression y M 1 M N N 1 Recovery: compute from y by M N N N arg min z 1 z s.t. y z 8/19

16 Compressive Sensing Basics = Compression y M 1 M N N 1 Recovery: compute from y by N N Representation basis (only used for recovery) arg min z 1 z s.t. y M N z 8/19

17 Compressive Sensing Basics = Compression y M Recovery: compute from y by N N For better recovery Ψ sparsify 1 M N N 1 Representation basis (only used for recovery) arg min z 1 z s.t. y M Ψ -1 has many zeros N z 8/19

18 Trace-Driven Design Select Φ and Ψ based on traces Data traces from 40 branches for 18 hours Classify power consumption patterns Duty-cycled (fridge) Periodic (projector) Time (seconds) Time (seconds) Fluctuating (desktop) Spiky (server) Time (seconds) Time (seconds) 9/19

19 Random Matri Φ Gaussian, Bernoulli, Binary recovery error ~ 2 2 ~ : : original recovered Recovery error (%) Representation basis differential transform cosine transform Haar wavelet transform 10/19

20 Random Matri Φ Gaussian, Bernoulli, Binary recovery error ~ 2 2 ~ : : original recovered Recovery error (%) Representation basis differential transform cosine transform Haar wavelet transform 10/19

21 Random Matri Φ Gaussian, Bernoulli, Binary recovery error ~ 2 2 ~ : : original recovered Recovery error (%) Representation basis differential transform cosine transform Haar wavelet transform 10/19

22 Random Matri Φ Gaussian, Bernoulli, Binary recovery error ~ 2 2 ~ : : original recovered Representation basis dominates recovery performance Use binary random matri Recovery error (%) Representation basis differential transform cosine transform Haar wavelet transform 10/19

23 Representation Basis Ψ Differential transform (Diff) Cosine transform (Cos) Haar wavelet transform (Haar) sparsity # of nonzeros signal length Average sparsity duty-cycled periodic fluctuating spiky 11/19

24 Representation Basis Ψ Differential transform (Diff) Cosine transform (Cos) Haar wavelet transform (Haar) sparsity # of nonzeros signal length Average sparsity duty-cycled periodic fluctuating spiky Best choice: Diff Cos Haar Diff 11/19

25 Representation Basis Ψ Differential transform (Diff) Cosine transform (Cos) Haar wavelet transform (Haar) sparsity # of signal nonzeros length Changing power pattern: TV A plug monitors multiple appliances Average sparsity Adapt Ψ to changing power pattern duty-cycled periodic fluctuating spiky Best choice: Diff Cos Haar Diff 11/19

26 Adaptive Representation Basis Machine learning approach Plug selects Ψ based on shape features Ψ = Diff Ψ = Cos Ψ = Haar Shape feature 1 Compressed signal Smart plug Choice of Ψ Base station 12/19

27 Adaptive Representation Basis Machine learning approach Plug selects Ψ based on shape features Base station learns decision boundaries Ψ = Diff Ψ = Haar Ψ = Cos learn and update to plug Shape feature 1 Compressed signal Smart plug Choice of Ψ Base station 12/19

28 Shape Feature & Decision Table shape feature vector = # of zero crossings # of sharp changes standard deviation # of zero crossings > Δ 1? N N N N Y Y Y Y # of sharp changes > Δ 2? N N Y Y N N Y Y Standard deviation > Δ 3? N Y N Y N Y N Y Choice of basis ADT ADT HWT DCT HWT HWT ADT DCT 13

29 Shape Feature & Decision Table shape feature vector = # of zero crossings # of sharp changes standard deviation # of zero crossings > Δ 1? N N N N Y Y Y Y # of sharp changes > Δ 2? N N Y Y N N Y Y Standard deviation > Δ 3? N Y N Y N Y N Y Choice of basis ADT ADT HWT DCT HWT HWT ADT DCT 13

30 Shape Feature & Decision Table shape feature vector = # of zero crossings # of sharp changes standard deviation # of zero crossings > Δ 1? N N N N Y Y Y Y # of sharp changes > Δ 2? N N Y Y N N Y Y Standard deviation > Δ 3? N Y N Y N Y N Y Choice of basis ADT ADT HWT DCT HWT HWT ADT DCT Trained at base station Minimize recovery error 13

31 Outline Motivation Design of JICE Secrecy of JICE Eperiment 14/19

32 Statistics Leak and Perturbation Φ is unknown to attacker Provide a computational guarantee of secrecy [Rachlin 2008] Leak l 2 -norm, mean and variance Zero mean: appliance is off High mean: appliance is on 15/19

33 Statistics Leak and Perturbation Φ is unknown to attacker Provide a computational guarantee of secrecy [Rachlin 2008] Leak l 2 -norm, mean and variance ~ n n [ k,0,0,,0] T 15/19

34 Statistics Leak and Perturbation Φ is unknown to attacker Provide a computational guarantee of secrecy [Rachlin 2008] Leak l 2 -norm, mean and variance ~ n Representation basis n [ k,0,0,,0] T 15/19

35 Statistics Leak and Perturbation Φ is unknown to attacker Provide a computational guarantee of secrecy [Rachlin 2008] Leak l 2 -norm, mean and variance ~ n Representation basis n [ k,0,0,,0] T shared secret 15/19

36 Statistics Leak and Perturbation Φ is unknown to attacker Provide a computational guarantee of secrecy [Rachlin 2008] Leak l 2 -norm, mean and variance ~ n Representation basis n [ k,0,0,,0] T Statistics depend on k shared secret 15/19

37 Statistics Leak and Perturbation Φ is unknown to attacker Provide a computational guarantee of secrecy [Rachlin 2008] Leak l 2 -norm, mean and variance ~ n n [ k,0,0,,0] T Statistics depend on k Little (no) change to sparsity Little impact on recovery Very sparse in transform domain 15/19

38 Recap of JICE Seed for generating Φ Key for generating n Smart plug Base station 16

39 Recap of JICE Seed for generating Φ Key for generating n Smart plug Base station power signal perturb & compress y = Φ ( + n) y 16

40 Recap of JICE Seed for generating Φ Key for generating n Smart plug Base station power signal Feature etraction perturb & compress y = Φ ( + n) Decision table Ψ choice (2 bits) y 16

41 Recap of JICE Seed for generating Φ Key for generating n Smart plug Base station power signal Feature etraction perturb & compress y = Φ ( + n) Decision table Ψ choice (2 bits) y recover & de-perturb = Ψ argmin z - n s.t. y = ΦΨz Recovered signals 16

42 Recap of JICE Seed for generating Φ Key for generating n Smart plug Base station power signal Feature etraction perturb & compress y = Φ ( + n) Decision table Ψ choice (2 bits) y updates (14 B) recover & de-perturb = Ψ argmin z - n s.t. y = ΦΨz Decision table training Recovered signals 16

43 Recap of JICE Seed for generating Φ Key for generating n Smart plug Base station power signal Feature etraction perturb & compress y = Φ ( + n) Decision table Ψ choice (2 bits) y updates (14 B) recover & de-perturb = Ψ argmin z - n s.t. y = ΦΨz Decision table training Recovered signals eecuted every a few hours 16

44 Outline Motivation Design of JICE Secrecy of JICE Eperiment 17/19

45 Implementation Smart plug [Sonnonet] Kmote Smart plug Kmote (8MHz MCU, 10KB RAM, ZigBee, TinyOS) Baselines Pipeline: Lossy compressor [Liu 2013] + AES Downsampling Lossless pipeline: SLZW + AES 18/19

46 Implementation Smart plug [Sonnonet] Kmote Smart plug Kmote (8MHz MCU, 10KB RAM, ZigBee, TinyOS) Same compression ratio with JICE Baselines Pipeline: Lossy compressor [Liu 2013] + AES Downsampling Lossless pipeline: SLZW + AES 18/19

47 Adaptive Basis vs. Fied Basis For TV Recovery error (%) Time 19/19

48 Adaptive Basis vs. Fied Basis For TV Recovery error (%) Time JICE achieves best performance with one eception 19/19

49 Adaptive Basis vs. Fied Basis For TV Recovery error (%) Time JICE achieves best performance with one eception 19/19

50 Data Fidelity and Scalability Pipeline Background traffic node JICE 20/19

51 Data Fidelity and Scalability Pipeline Background traffic node JICE = 12 X 20/19

52 Data Fidelity and Scalability Pipeline JICE Pipeline # of background traffic nodes Background traffic node JICE = 12 X 20/19

53 Data Fidelity and Scalability Pipeline JICE Project to 96 plugs Project to 144 plugs Pipeline # of background traffic nodes JICE supports 50% more plugs Background traffic node = 12 X JICE 20/19

54 Conclusion & Future work JICE Supports more nodes for same data fidelity Better data secrecy than pure compressive sensing Adaptive to changing power pattern Future work Other applications 21/19