1

Signature-based Detection for Activities ofAppliances

Zhichuan Huang, Ting Zhu and Hongyao LuoUniversity of Maryland Baltimore County, Baltimore, USA

Abstract—The large-scale placement of smart meters bringsconcerns about privacy of individuals in their own homes. Energyconsumption data can reveal precise appliances’ usage informa-tion with Non-Intrusive Load Monitoring (NILM). To mitigatebehavior leakage in homes with smart meter measurement,Battery-based Load Hiding (BLH) is widely studied. In thisapproach, a rechargeable battery is used to store and supplypower to home appliances at strategic time to hide appliances’consumption. However, because different appliances have differ-ent energy consumption signatures, it is still possible to detectactivities of appliances. In this paper, we developed signature-based detection (SBD) method, which utilizes the signatures ofappliances to detect appliances usage. We conducted extensivesystem evaluations with 40 homes. Results indicate that i) ourconsumption signature detection method has higher detectionaccuracy than edge detection; ii) our consumption detectionmethod can still detect more than 70% of appliances’ eventseven with BLH.

I. INTRODUCTION

Smart meters are deployed in homes to record high granular-ity energy consumption for power grid monitoring. However,the large-scale placement of smart meters has introducedpotential leakage of private and valuable information aboutoccupants’ activities [1]. Moreover, many utility companies areproviding third-party companies access to traces of smart meterdata, which results in high probability of privacy leakage. Withthe technology of Non-Intrusive Load Monitoring (NILM) [2],the high granularity energy usage data obtained can be ana-lyzed to reveal personal information such as usage of individ-ual appliances, sleep patterns, number of occupants and evenoccupants’ health conditions. The widely used technique is theedge detection [3], which looks for the sharp edges that revealthe significant changes in the steady current consumed by thehousehold. To prevent private information leakage, researcherspropose Battery-based Load Hiding (BLH) algorithms [4] [5],which employ batteries to partially supply the net demand loadfrom the house to alter the external load as seen by the smartmeter. The battery is charged and discharged at specific time tohide the energy consumption events. However, BLH algorithmshave to cope with limited battery capacity and discharge rates.Current methods also require users to charge and dischargebatteries frequently, which will significantly decrease the bat-tery’s lifetime. Furthermore, batteries will introduce the energyconversion loss during charging and discharging stages.

Recently, researchers revealed the empirical power con-sumption signatures of different electrical loads [6]. The resultsshow that the power consumption of electrical loads is not onlyeither zero or stable power consumption. During the usage,

the power consumption of electrical loads varies significantlyat different stages while follows a unique signature. Thesignatures of electrical loads provide great opportunity forappliances detection. We developed a new consumption detec-tion method, called signature-based detection (SBD) method,which reveals appliances’ usage more accurately than existedapproaches (e.g., edge detection). SBD can be used by utilitycompanies to provide energy-efficiency recommendations forconsumers. However, by taking advantage of our method, themalicious third parties can also easily detect appliances’ usageif they have empirical power consumption data. This will resultin a serious private information leakage.

To evaluate our proposed SBD method and compare it withexisting ones, we conduct real-world experiments by deployingenergy meters in multiple homes to get the energy consumptionsignatures of individual appliances. We also run large-scalesimulations by using the empirical energy consumption tracesfrom 40 homes. Our simulation results indicate i) SBD methodhas higher detection accuracy than edge detection; ii) SBDmethod can still detect more than 70% of appliances’ eventseven with BLH.

The rest of the paper is organized as follows: backgroundof edge detection, BLH and appliance consumption signaturesare introduced in §II; the detailed design of SBD is describedin §III; implementation and simulations are provided in §IV;finally, we conclude the paper in §V.

II. BACKGROUND

In this section, we i) state current problem of privacyexposure in sensing system; ii) introduce the widely used edgedetection method; iii) demonstrate the mechanism of BLH;and iv) describe the recently developed energy consumptionsignature models of different appliances.

A. Privacy in Sensing Systems

With the large deployment of different types of sensors,privacy issue of sensing systems has become an importantproblem [7]. In [8], a new theoretical scheme called SensorSiftis introduced for balancing utility and privacy in smart sensorapplications. In [9], a theoretical framework is proposed toallow users to quantify the utility-privacy tradeoff in smartmeter data. In [10], authors explore three case studies andanalysis of secure cyber-physical systems: wireless medicaldevices, robots, and automobiles.

More seriously, many utilities are providing third-partycompanies access to traces of smart meter data. Third-party

2

companies are now employing cloud-based platforms to an-alyze large amount of smart meter data [12] [11]. The datacollected from smart meters can be applied in different energymanagement applications [13] [14]. While the purpose is toprovide consumers energy-efficiency recommendations, com-panies are also mining the data for profitable information suchas usage of certain brand of appliances, which can be usedfor future advertisement. Some people may argue that there isno need to hide those information. However, appliances usageinformation can be used to reveal more private information.For example, usage time of certain appliances (e.g., waterheater) can reveal the number of people that live in the house.Besides that, changes of appliances’ usage patterns can alsoindicate other personal information. For example, if your waterdispenser only works at evening when you come back home,and suddenly your water dispenser works frequently in day-time and other appliances’ usage pattern stays the same; wecan guess that you may be sick at home and need to drink a lotof water. As a result, it is critical to protect energy consumptionprivacy for individuals.

B. Edge Detection Method

NILM algorithms have been widely used in research ofresidential settings to reveal usage of individual applianceswith consumption data [15]. Moreover, many applications havebeen proposed to extend existing NILM algorithms. In [1],appliances’ usage patterns are possible to be extracted frompower consumption data without prior knowledge of householdactivities and training. In [3], NILM algorithms are extended toevaluate the threat to individual privacy by considering resultson potential disclosure from smart meter data. Most of theseNILM algorithms are based on edge detection.

Edge detection technique aims to looks for significantchanges in the energy being consumed by the household[5]. Such changes are characterized by sharp edges in theenergy consumed by the appliances. These edges are thenclustered and matched against known appliance profiles. Forinstance, if someone turns on a 20W lamp, then the net powerconsumption increases by 20W . Conversely, when the lampis turned off, the net current drops by the same amount. Thealgorithm detects the pair of edges with equal magnitude andopposite direction, and match them against the electric profiles.

C. Battery-based Load Hiding

To avoid edge detection, the most intuitive way is tomake sure that power consumption recorded at smart meterside is much higher or lower than real demand, then edgedetection method would fail to detect the real appliance thatis turned on. Let current power demand of one appliancebe d(t). To disable edge detection, we need only to findthe appliance’s modeled data mi(t) < d(t) that minimizesd(t)mi(t) and appliance’s modeled data mi+1(t) > d(t)that minimizes mi+1(t)d(t). Then we have minimum powerchange ζei (t) = [mi+1(t)d(t)]/2 and δei (t) = [d(t)mi(t)]/2.For example, given that power consumption of a lamp is 40Wand power consumption of laptop is 100W , then, if we turn on

the lamp, meanwhile increasing power consumption from 40Wto 100W, edge detection method would detect that the workingappliance is the laptop instead of the lamp. One widely appliedapproach to increase or decrease power consumption is BLH.

The basic idea of BLH is to use a rechargeable battery tostore and supply power to home appliances at strategic timesto hide appliance’s consumption from smart meters [16]. Thereare already many proposed BLH algorithms. The Best Effort(BE) algorithm [17] tries to avoid charging the external loadwhenever possible, and when the actual demand is differentfrom external load, the battery can be charged or dischargedto counteract the difference. The Non-Intrusive Load Leveling(NILL) algorithm [4] is proposed to have three states of loadhiding. Within each state, the NILL maintains a differentconstant load to hide real appliances’ usage. However, newpitfalls are also identified in these two aforementioned BLHalgorithms (BE and NILL), which allow for the recoveryof precise appliance usage information [5]. To address suchproblem, a novel stepping-based module for BLH algorithmsis developed to defend against precise load change recoveryattacks. However, there are still some actual cases that BLHapproaches cannot address perfectly due to the battery’s phys-ical limitations (e.g., limited capacity and slow charging anddischarging speed).

D. Power Consumption SignatureRecently, researchers revealed the empirical power con-

sumption signature of different electrical loads [6]. They clas-sified electrical loads into five consumption signature models:On-Off Model. An on-off model includes two states: an onstate that draws some fixed power pon and an off state thatdraws zero, or some minimal amount of power poff . A goodexample is lamp and its energy consumption pattern can beformulated as [6]:

p(t) =

{pon 0 ≤ t ≤ tonpoff t ≥ ton

(1)

On-Off Growth/Decay Model. An on-off growth/decay modelis a variant of the on-off model that accounts for an initialpower surge mpeak when a load starts, followed by a smoothincrease or decrease in power usage over time until it reachesa stable consumption mon. An example of this model isan air conditioner (AC) and its consumption pattern can beformulated as [6]:

p(t) =

{pon + (ppeak − pon)e−λt 0 ≤ t ≤ tonpoff t ≥ ton

(2)

Stable Min-Max Model. A stable min-max model maintainsa stable maximum (mmax) or minimum power consumptionwhen in active state, but frequently with a spike (mspike)from this stable state. r(t) can only be 0 or 1. For example,refrigerator works with minimum power at most of time andperiodically wake up to keep temperature low with high powerconsumption. The consumption pattern can be formulatedas [6]:

p(t) = pmax · r(t), 0 ≤ r(t) ≤ 1 (3)

3

Random Range Model. A random range model draws a ran-dom amount of power within a fixed range (mmin to mmax).Television is an example of Random Range model becauseits energy consumption depends on display content and variesfrom a minimum and maximum power. The consumptionpattern can be formulated as [6]:

p(t) = pmin + (pmax − pmin) · r(t), 0 ≤ r(t) ≤ 1 (4)

Composite Model. An composite model exhibits characteris-tics of multiple basic model types either in sequence or par-allel. For example, a washing machine with different workingstages may consume different energy signature at differentworking stages.

During the usage, the power consumption of electricalloads varies significantly at different stages. The signaturesof electrical loads provide great opportunity for appliancesdetection.

III. SIGNATURE-BASED DETECTION METHOD

While edge detection methods are simple, they are ofteninaccurate, because they fail to capture the complex powerusage patterns of different loads. With the energy consumptionsignatures of different appliances described above, we can de-sign a better method than edge detection to reveal appliances’usage patterns with a home’s energy consumption data. Thekey idea is to detect appliances usage by the similarity betweenreal power consumption and appliances’ consumption models.If consumption model of an appliance ai is most similar to realpower consumption, then it is highly possible that appliance aiis working but not other appliances. In this paper, we proposea Euclidean distance-based function to quantify the similaritybetween two vectors. Let e(t) be the real energy consumptionand mi(t) be energy consumption data generated by models attime t, where length(ai) is the signature length of applianceai. The similarity between two vectors can be calculated as:

ρi =1

1 + l(e,mi)(5)

where

l(e,mi) =1

length(ai)

length(ai)∑t=1

(e(t)−mi(t))2 (6)

Equation (6) is used to calculate the distance between twovectors. Because different appliances’ models have differentlengths of signature sequences, we use 1/T to normalize thedistance of two vectors. For example, signature sequences ofa lamp is short due to the on-off model; while the signaturesequences of TV is long due to dynamic power consumptionduring usage. Equation (5) is used to transfer distance tosimilarity with range of [0, 1].

Based on the similarity between consumption models ofdifferent appliances and real consumption data, we detect theappliances’ usage patterns. Suppose appliance i has highestsimilarity with real power consumption from time t, we thendetect as appliance i is working. Because several appliancescan be working at the same time, we can remove the detected

Algorithm 1: Signature-Based Detection AlgorithmInput: Home’s power consumption pOutput: Appliances usage.

1: t=1;2: while t < T do3: ρmax = 0, s = −1;4: for Appliance ai do5: Calculate ρi(t) based on Equation 5;6: if ρi(t) > ρmax then7: ρmax = ρi(t), s = i,8: end if9: end for

10: Calculate ρmin;11: if ρmax > ρmin then12: Detect appliance ai working at t;13: p(t : t+ length(as)) = p(t :

t+ length(as))−mi(1 : length(as));14: else15: t = t+ length(as);16: end if17: end while

appliance’s model from real data and then repeat the detectionprocess again. When similarity between rest of appliances andreal consumption is low, we end the detection process fortime t and continue the detection process from the time whendetected appliances stop working.

The details of detection algorithm is shown in Algorithm1. For power consumption data from t = 1, · · · , T , wefirst initialize t = 1 (Line 1). While t < T , we calculatesimilarity between power consumption data and signatures ofeach appliances based on Equation (5) (Lines 2-5). If wefind the similarity between power consumption and signaturesof an appliance is higher than current maximum similarity,we reassign maximum similarity and mark s = i (Lines 6-9). Then we calculate similarity between power consumptiondata and zero power consumption to get minimum similarityρmin(Line 10). If ρmax > ρmin, we then detect appliance asis working at t and update power consumption by removingthe signature of appliance as (Lines 11-13). Note t stays thesame value to continue detecting other appliances working attime t. Otherwise, it means we already detect all the appliancesworking from t, thus we update t = t+ length(as) for furtherdetection (Lines 14-17).

IV. IMPLEMENTATION AND EVALUATION

In this section, we evaluate the performance of our signaturebased detection method and compare it with edge detectionmethod.

A. Data CollectionWe deploy eGauge power meters at individual homes to

collect the total energy consumption data every one second.One of the experiment setup is shown in Figure 1(a). In oursimulation, we use the energy consumption traces collected

4

(a) Experiment setup

06/09 06/10 06/11 06/12 06/13 06/14 06/150

4

8

12

16

Time (seconds)

Pow

er (

kW)

(b) Energy consumption of one home in six days

(c) Battery (d) Inverter (e) Battery Measurement

Fig. 1. Experiment setup and data collection

0 300 600 900 1200 15000

1

2

3

4

Time (s)

Pow

er

(kW

) Microwave

Oven

Refridgerator

Refridgerator

Trash

Compactor

AC

Coffee Maker

Oven

Refridgerator

TV

Fig. 2. An example of detection results

from 40 homes (shown in Figure 1(b)). We also collectload events of one home to get consumption signature ofall the electrical loads (e.g., TV, oven, etc.). Then withpower consumption signature, other homes’ load events aredetected as ground truth. The energy storage unit we deploy isUB12100-S Universal Battery and Xantrex PowerHub 84053shown in Figure 1(c) and 1(d), which is a combination of aninverter/charger module capable of delivering up to 1800 wattsof household power. We use iMeter Solo (an INSTEON powermeter) to measure the battery energy charging and dischargingrate in real time (shown in Figure 1(e)).

B. Evaluation ResultsWith empirical consumption and load events data collected

in 40 homes, we run both edge detection and SBD methods. Anexample of detection results at one home is shown in Figure 2.Microwave oven belongs to On-Off model; refrigerator andtrash compactor belong to Stable Min-Max model; oven andcoffee maker belong to Composite model; AC belongs to On-Off Decay model; TV belongs to Random Range model. Wealso show whether appliances can be detected by using onlyedge detection or SBD methods. Because the microwave ovenand coffee maker have similar power consumption, they cannot

Appliance Model Edge SignatureMicrowave Oven On/Off No Yes

Refrigerator Stable Min-Max Yes YesOven Composite Yes Yes

Trash Compactor Stable Min-Max No YesCoffee Maker Composite No Yes

AC On/Off Decay Yes YesTV Random Range No Yes

TABLE I. DETECTION RESULTS

1 2 3 4 5 6 760%

70%

80%

90%

100%

Number of appliancesAp

pli

ance

s d

etec

tio

n a

ccu

racy

Edge detection

SBD

Fig. 3. Detection accuracy with original power consumption

be detected by edge detection. However, the coffee makerhas two working stages, which can be differentiated from themicrowave oven by using SBD. Trash compactor and TV alsohave similar power consumption but different working stages,thus they cannot be differentiated by edge detection but signa-ture detection. The summary of different appliances detectionis shown in Table I. Therefore, our signature detection methodcan detect appliances’ usage more accurately.

We also evaluate appliances detection accuracy for a singlehome in 30 days. The results are shown in Figure 3. We definedetection accuracy to verify the performance of our design. Ingiven time period, the number of state changes (on to off oroff to on) of appliances is Nt and the number of correctlydetected state changes is Nd. Then the detection accuracy rcan be calculated as:

r =NdNt

(7)

In the simulation, we use collected appliances events andappliances’ real-time power consumption to generate totalpower consumption to detect. With original power consump-tion data, both edge detection and SBD can detect appliancesusage with more than 70% accuracy. With more appliances todetect, the detection accuracy of edge detection and SBD bothdrops. This is because different appliances may have similarpower consumption signature, which increases the difficulty todetect more appliances. However, compared to edge detection,detection accuracy of SBD drops much slower.

We also evaluate appliances detection accuracy when powerconsumption is hidden by BLH. The power consumption forcharging a battery is shown in Figure 4. The average powerfor charging the battery is around 160W. The BLH algorithmwe choose is from [5]. Yang et al. proposed four battery-based algorithms to hide power consumption. In our paper,we select (Lazy Stepping) LS2 because LS2 performs best inmost of their simulations. We show the power consumption

5

0 20 40 60 80 100 120 140 160 180150

160

170

180

190P

ow

er (

W)

Time (Minute)0 20 40 60 80 100 120 140 160 180

0.5

0.6

0.7

0.8

0.9

1

Ch

arg

ing

eff

icie

ncy

Fig. 4. Battery Charging

0 60 120 180 240 300−2

0

2

4

6

Time (s)

Po

wer

(kW

)

LS2

Original Load

Fig. 5. Original load and hidden load with LS2

of BLH algorithm LS2 with comparison of the original loadsin Figure 5. To make the difference between original loadand power consumption with BLH visible, we show only300 seconds of consumption data in one home. The LS2tries to maintain power consumption at certain levels, thus itsconsumption can be only -2kW , 0kW , 2kW , 4kW and 6kW .However, we can still find that the shape of LS2 is similar tothe original load. The results of two detection algorithms areshown in Figure 6. For edge detection, the detection accuracyof 7 appliances drops below 50% while the detection accuracyfor SBD is still higher than 70%.

With our proposed SBD method, appliances’ usage canbe detected accurately even power consumption is hidden byBLH. Thus, there will be stronger demand for homeowners toprotect their appliances’ usage.

1 2 3 4 5 6 740%

60%

80%

100%

Number of appliancesAp

pli

ance

s d

etec

tio

n a

ccu

racy

Edge detection

SBD

Fig. 6. Detection accuracy with original power consumption hidden by BLH

V. CONCLUSION

In this paper, we demonstrate the recently proposed energyconsumption signature models and develop our energy con-sumption signature detection method, called SBD, which ismore accurate than the widely used edge detection method.With the empirical data from more than 40 homes, we conductextensive system evaluations. Results indicate that i) SBD hashigher detection accuracy than edge detection; ii) SBD can stilldetect more than 70% of appliances’ events even with BLH.

ACKNOWLEDGMENT

This work was supported by NSF CNS-1503590.

REFERENCES

[1] A. Molina-Markham, P. Shenoy, K. Fu, E. Cecchet, and D. Irwin,“Private memoirs of a smart meter,” in BuildSys, 2010.

[2] G. Hart, “Residential energy monitoring and computerized surveillancevia utility power flows,” IEEE Technology and Society Magazine, vol. 8,no. 2, pp. 12–16, 1989.

[3] M. Lisovich, D. Mulligan, and S. Wicker, “Inferring personal informa-tion from demand-response systems,” IEEE Security Privacy, vol. 8,no. 1, pp. 11–20, 2010.

[4] S. McLaughlin, P. McDaniel, and W. Aiello, “Protecting consumerprivacy from electric load monitoring,” in CCS, 2011.

[5] W. Yang, N. Li, Y. Qi, W. Qardaji, S. McLaughlin, and P. McDaniel,“Minimizing private data disclosures in the smart grid,” in CCS, 2012.

[6] S. Barker, S. Kalra, D. Irwin, and P. Shenoy, “Empirical characterizationand modeling of electrical loads in smart homes,” in IGCC, 2013.

[7] M. Ali, E. Al-Shaer, and Q. Duan, “Randomizing ami configuration forproactive defense in smart grid,” in SmartGridComm, 2013.

[8] M. Enev, J. Jung, L. Bo, X. Ren, and T. Kohno, “Sensorsift: balancingsensor data privacy and utility in automated face understanding,”in Proceedings of the 28th Annual Computer Security ApplicationsConference, 2012.

[9] L. Sankar, S. Rajagopalan, S. Mohajer, and H. Poor, “Smart meterprivacy: A theoretical framework,” IEEE Transactions on Smart Grid,vol. 4, no. 2, pp. 837–846, 2013.

[10] T. Kohno, “Security for cyber-physical systems: case studies withmedical devices, robots, and automobiles,” in Proceedings of the fifthACM conference on Security and Privacy in Wireless and MobileNetworks, 2012.

[11] R. P. Singh, S. Keshav, and T. Brecht, “A cloud-based consumer-centricarchitecture for energy data analytics,” in e-Energy, 2013.

[12] Z. Huang, H. Luo, D. Skoda, T. Zhu, and Y. Gu, “E-sketch: Gatheringlarge-scale energy consumption data based on consumption patterns,”in IEEE International Conference on Big Data, pp. 656–665, IEEE,2014.

[13] T. Zhu, Z. Huang, A. Sharma, J. Su, D. Irwin, A. Mishra, D. Menasche,and P. Shenoy, “Sharing renewable energy in smart microgrids,” inICCPS, pp. 219–228, ACM, 2013.

[14] Z. Huang, T. Zhu, Y. Gu, D. Irwin, A. Mishra, and P. Shenoy, “Min-imizing electricity costs by sharing energy in sustainable microgrids,”in BuildSys, pp. 120–129, ACM, 2014.

[15] A. Marques, M. Serrano, S. Karnouskos, P. J. Marron, R. Sauter, E.Bekiaris, E. Kesidou and J. Hoglund, “Nobel - a neighborhood orientedbrokerage electricity and monitoring system,” in e-Energy, 2010.

[16] A. Rial and G. Danezis., “Privacy-preserving smart metering,” inTechnical Report MSR-TR-2010-150, Microsoft Research, 2010.

[17] T. Carpenter, S. Singla, P. Azimzadeh, and S. Keshav, “The impact ofelectricity pricing schemes on storage adoption in ontario,” in e-Energy,2012.