Differentially Private Demand Side Management for Incentivized Dynamic Pricing in Smart Grid
aa r X i v : . [ c s . CR ] F e b Differentially Private Demand Side Management forIncentivized Dynamic Pricing in Smart Grid Muneeb Ul Hassan ∗ , Mubashir Husain Rehmani § , Jinjun Chen ∗∗ Swinburne University of Technology, Hawthorn VIC 3122, Australia § Munster Technological University (MTU), Ireland
Abstract —In order to efficiently provide demand side manage-ment (DSM) in smart grid, carrying out pricing on the basis ofreal-time energy usage is considered to be the most vital toolbecause it is directly linked with the finances associated withsmart meters. Hence, every smart meter user wants to pay min-imum possible amount along with getting maximum benefits. Inhere, usage based dynamic pricing strategies of DSM plays theirrole and provide users with specific incentives that help shapingtheir load curve according to the forecasted load. However,these reported real-time values can leak privacy of smart meterusers, which can lead to serious consequences such as spying,etc. Moreover, most of dynamic pricing algorithms charges allusers equally irrespective of their contribution in causing peakfactor. Therefore, in this paper, we propose a modified usagebased dynamic pricing mechanism that only charges the usersresponsible for causing peak factor. We further integrate theconcept of differential privacy to protect the privacy of real-time smart metering data and to calculate accurate billing, wepropose a noise adjustment method. Finally, we propose DemandResponse enhancing Differential Pricing (DRDP) strategy thateffectively enhances demand response along with providing dy-namic pricing to smart meter users. We also carry out extensivetheoretical analysis for differential privacy guarantees and forcooperative state probability to analyse behaviour of cooperativesmart meters. The performance evaluation of DRDP strategyat various privacy parameters show that the proposed strategyoutperforms previous mechanisms in terms of dynamic pricingand privacy preservation.
Index Terms —Differential Privacy (DP), Smart Grid (SG),Demand Side Management (DSM), Dynamic Pricing, PrivacyPreservation, Demand Response (DR).
I. I
NTRODUCTION
Modern day smart homes are equipped with smart meterswhich send their their real-time energy usage values tosmart grid utility in order to carry out plenty of taskssuch as demand side management (DSM), load forecasting,etc [1]. This real-time energy usage data is used to formulizestrategies that help shape the load curves and carry outefficient load utilization (ELU). ELU is a method of shapingsmart homes energy usage in a such a way that it equates withthe possible energy supply in the specific time instant [2]. Inorder to do so, demand side management (DSM) strategiescame into discussion, which shape the load curves byproviding interesting and timely incentives to participating A preliminary version has been published by 2020 IEEE InternationalConference on Communications (ICC 2020), June, 2020, Dublin, Ireland en-titled “Differentially Private Dynamic Pricing for Efficient Demand Responsein Smart Grid”.This paper is partly supported by Australian Research Council (ARC) projectsDP190101893, DP170100136, LP180100758. smart homes [3]. Similarly, almost all DSM strategies (alsoknown as demand response (DR) strategies) have a commongoal, which is to motivate smart homes users to use minimumenergy during peak load times and to shift surplus energyusage to off-peak times (e.g., using washing machine inoff-peak hours) [4].Till now, plenty of DR models have been discussed inresearch, for example, control mechanism models, offeredmotivation models, and decision variable models. Amongthem, offered motivation models are the most popular oneswhich are further categorized into price based models andincentive based models [5]. In offered motivation models,dynamic pricing dominate other models because it provideusers the maximum control to get incentivized. In dynamicpricing mechanisms, users are charged with respect to therate being devised by grid utility, so users can orient theirusage at the time of low rates and use heavy appliances atoff-peak hours. This model is somehow beneficial, but it hasa major flaw that what if all smart homes start using theirheavy appliances at once at the time of low pricing hours?
In fact, if this happens, then the low-pricing hours can getshortage of electricity as it was not predicted during loadforecasting and strategy to overcome this sudden shortagewas not developed. In order to overcome this, researcherscame up with the idea of dynamic peak hours, which meansthat peaks hours are not fixed and can vary with respect toenergy usage within a specific area. For example, if energyexceeds a specific peak value, then the peak-hour is in placeand smart homes will be charged peak hour price [3].Overall, this dynamic peak factor model is well-suited to meetthe demands of load forecasting, but on the other hand, it hastwo major issues from perspective of smart homes. Firstly,it do also charges the same high peak factor price to smarthomes which are not responsible to cause that peak-hour.Secondly, the collection of fine-grained data of smart homesfor load forecasting and for peak hour determination raisesserious threats to privacy leakage of smart home users. E.g.,this real-time data can further be used to carry out variousmalicious activities such as forgery, routine tracking, etc.Similarly, this data can also be fed-up to non-intrusive loadmonitoring (NILM) mechanisms, these techniques predictthe usage of a specific household appliance (such as toaster,washing machine, etc.) at a specific slot of time [6]. TheseNILM mechanisms can even find out any faulty appliance andcan estimate its possible day of breakdown, which can furtherbe used to carry out targeted advertisement [7]. Therefore, amechanism that provides both; usage based dynamic pricing along side preserving privacy of smart homes is required.In this paper, we first work over developing a dynamicpricing strategy that facilitates the participating users andonly charge the users who are responsible to cause thatpeak factor. In order to do so, we work over carrying outprivate data analysis that effectively tracks that whether theuser is responsible for peak factor or not. Furthermore, toensure privacy in the proposed strategy, we use the notionof differential privacy that adds independent and identicallydistributed (i.i.d) noise in the real-time metering values topreserve the privacy. The noise is added in such a manner thatthe data is still useful for billing, DSM, or load-forecasting.In this regard, we proposed a noise adjustment method tomaximize utility alongside preserving privacy. However, it isensured that NILM techniques will not be able to analyse theexact usage/appliance pattern due to added noise. Collectively,we propose D emand R esponse enhancing D ifferential P ricing(DRDP) mechanism that is responsible for both; private datareporting and usage based dynamic pricing. Experimentalevaluation of our proposed DRDP mechanism shows that ourmechanism incentivizes participating users by only chargingthe peak price to the users responsible for causing peak valuealong with providing the benefit of private reporting to smartgrid utility.The remainder of our paper is organized as follows; section2 provides discussion about previous literature review andother state-of-the-art works in similar field, section 3 providesdetailed discussion about system model, adversary model, andproblem formulation, section 4 provides comprehensive dis-cussion about proposed DRDP mechanism and its algorithmicfoundation, section 5 covers all aspects of performance eval-uation of DRDP, after that, the article is concluded in section6 by providing concluding remarks and future directions.II. L ITERATURE R EVIEW
Dynamic pricing is a well-researched domain and manyworks from perspective of dynamic pricing for smart grid havebeen carried out by researchers. However, the issue of privacypreservation while providing dynamic pricing has not beenwell addressed in plenty of works [14]. For instance, real-timedata from smart meters can cause privacy leakage becausethis data can further be used to analyse smart home users’behavioural pattern [15]. In order to protect this informationsome privacy preserving smart metering mechanisms havebeen developed in the literature that we will cover in thissection. Similarly, another major issue in dynamic billing isthat it charge all customers irrespective of their contributionin causing of peak factor. For example, even if a user is usingminimum amount of energy it charges that user the sametariff that the mechanism will charge to the person utilizingmaximum capacity of its load, which can even be termed asunjust billing [16]. In this section we cover the current state-of-the-art research that has been carried out over the topic tillnow.In current literature, certain works highlight the use of dynamicpricing in usage based scenario, for example, the most promi-nent work in this domain has been carried out by Liang et al. in [8]. In this work, authors worked over usage baseddynamic pricing and presented a model which uses distributedcommunity gateway for aggregation and price control fea-ture. In order to enhance privacy, authors used homomorphicencryption based privacy. The presented results demonstratethat the carried work enhances previous pricing models alongwith overcoming a certain type of privacy violation attacks.Similarly, another work in the field of dynamic billing hasbeen carried out by authors in [9]. The authors worked overoptimal dynamic pricing that ensures profitability to smartgrid utility. Authors specifically worked over the trade-offbetween utility and profit from operators’ perspective, whichis further managed by using demand supply theory. However,authors did not specifically integrated any privacy preservationmechanism nor focused over privacy at all. Another workin the similar domain has been carried out by Gope andSikdar in [3]. Authors proposed a data aggregation mechanismwhich also provided the benefit of dynamic pricing along withenhancing demand response in modern energy systems. Toadd it further, authors protected the privacy of this data byusing encryption on group based data aggregation. Experimen-tal results demonstrate that authors enhanced computationalcost to provide light-weight data aggregation and billing forsmart grid. Another work discussing dynamic pricing underthresholding policies have been carried out by authors in [10].Authors developed two optimal dynamic pricing mechanismsand made greedy and sliding window heuristics for dynamicpricing. The proposed mechanism ensured that the pricingis adjusted according to power demand and supply. Authorsfurther enhanced approximation ratio and execution time oftraditional dynamic billing approaches.The other direction in literature review is the integration ofprivacy preservation especially differential privacy in real-timereporting to protect smart home users’ privacy. In order to doso, a solid work has been presented by Won et al. in [11].Authors work over integration of the concept of differentialprivacy with a fault-tolerant grid communication network,which can provide both benefit of privacy and fault tolerance.Authors carried out this work by using future ciphertexts andutilized their benefits to formulate a protocol that providescommunication efficiency along with providing reduction inroot mean square error. Another work in this field has beencarried out by Ni et al. that also covers the similar domain ofintegration of differential privacy with fault tolerance mech-anism [12]. Furthermore, this proposed mechanism do alsointegrate range-based filtering and EIGamal encryption forsecure and filtered output. The proposed mechanism of authorsenhanced communication overhead along with providing effi-cient reduction in computational costs. A work that discussedintegration of differential privacy for smart meters with re-newable energy resources (RER) for real-time smart meteringhas been presented by authors in [7]. The work covers theaspect of real-time load monitoring along with protecting userload profiles. The provided work overcomes eavesdroppingattacks and differential attack by integrating strong notion ofdifferential privacy with RER data. In this way, authors areable to protect both: smart metering usage privacy and RERgeneration and consumption privacy for microgrid users. A
TABLE I
A T
HOROUGH A NALYSIS OF D YNAMIC B ILLING AND P RIVATE S MART M ETERING M ECHANISM I N E NERGY S YSTEMS . MajorCategory RefNo. Focus ofArticle Mechanism Type MechanismFunctioning PrivacyType Metrics Enhanced TackledAttacks SimulationPlatform Compl-exity [8] Dynamicbilling UDP: Usage baseddynamic pricing Price control &aggregation viadistributed communitygateway & price control Homomo-rphicencryption • Pricing Model • Privacyviolation attack − O ( n/ DynamicPricing [9] Profit &utilitytrade-off Profitable smartgrid by developingoptimal dynamicpricing Managed trade-offbetween users’ utilityand operator profit viademand supply theory − • Average profit &utility• Efficiency ofPower utilization − AMPL − [3] Dynamicuserbilling Efficient dynamicbilling & demandresponse Using key encryptionto protect dataconfidentiallity forgroup based usage dataaggregation Encryption • Computationalcost • Eavesdrop-ping attack• DoS attack JCE,JPBE − [10] Dynamicpricingunderthresh-oldingpolicies Developed twooptimal dynamicpricing mechanisms Greedy & Sliding-Window heuristicfor price developedaccording to powerdemand − • ApproximationRatio• Execution Time • − JavaCPLEX
Multiple [11] Privateaggrega-tion Fault-tolerant &privacy assured smartgrid aggregation Worked over integrationof differential privacywith fault tolerantprotocol Differentialprivacy • Efficiency incommunication• RMSE • Curiousaggregationattack Realisticenergytraces O (log N + 1) PrivateGrid Re-porting [12] Data ag-gregation Using range basedfiltering to carryout private smartmetering Integrated range basedfiltering, differentialprivacy, & EIGamalencryption Differentialprivacy &encryption • Computationaloverhead• Communicationoverhead • False datainjection• Differentialattack Python,Miracl − [7] Protectingpeakdata andRERdata DPLM: DifferentiallyPrivate usagemonitoring withRER Integrated differentialprivacy with propertiesof intermittent nature ofRERs to preserve real-time reporting privacy Differentialprivacy • Load usageprofiles • Eavesdrop-ping attacks• Differentialattacks Python O ( n ) [13] Privacypre-servingattesta-tion AnonymousAttestation for SmartMeters Privacy protection &leakage detection viasignature in smartmeters AttributeCertificate& RingSignature • Computationaltime• Unforgetability • AdaptiveChosen-Plaintextattack Real-TimePrototype t + 2 PrivateDynamicBilling
ThisWork IncetivizedprivatedynamicBillingmecha-nism DRDP: Differeniallyprivate PrivateBilling with Usagebased Pricing Differential privacyprotection forsmart homes alongwith incentivizingcooperative users viadynamic pricing Differentialprivacy • Network trust• Privacy• Usage basedbilling • Mutualinformationsharing• Static &dynamicreporting Python O ( n ) work targeting privacy preserving attestation in smart meteringhave been carried out by authors in [13]. Authors developed ananonymous attestation mechanism for smart meters in whichthey protected privacy leakage via attribute certificates andring signatures. Authors enhanced computational complexityalong with working over unforgettability. A table for detailedcomparative analysis of all the mentioned mechanisms havebeen given in Table I. After carrying out careful analysis of all the previous works,it can summarized that to the best of our knowledge, no workthat integrates notion of dynamic differential privacy with real-time dynamic billing have been carried out in the literature.Similarly, in the preliminary work of this article [17], weanalyse the aspect of dynamic pricing and differential privacyon real-time smart metering data. In this extension article,we further propose a noise balancing mechanism for privatebilling, which can serve as first step forward in the directionof incentivizing users and enhancing demand response alongwith providing them strong privacy guarantee via differentialprivacy. III. P
ROVIDING D IFFERENTIALLY P RIVATE D YNAMIC B ILLING
In this section we demonstrate the motivation, problemformulation, system model functioning, and adversary modelfor our proposed DRDP mechanism.
A. Motivation of DRDP
The motivation for the proposed DRDP mechanism is givenbelow: • Conventional dynamic pricing mechanisms does not in-centive cooperative users and charges the same price toall users within a specific area. We propose a dynamicbilling strategy that only charges the users responsible tocause peak factor. • Traditional dynamic billing strategies does not incorpo-rate the notion of differential privacy to preserve privacyduring dynamic billing. However, in our DRDP strategy,we modified the approach of dynamic billing and inte-grated differential privacy as a privacy preserving notion.
B. Problem Formulation
We divide the problem formulation of our proposed DRDPmechanism into two parts first we discuss the privacy re-
Fig. 1:
The proposed system model for DRDP pricing where each smartmeter node send their differentially private readings to grid utility whichfurther adjusts the noise value via differential privacy adjustment for accuratebilling. quirements for dynamic billing and then we proposed threequestions that summarizes the problem formulation of ourDRDP model.
1) Privacy Requirements for Dynamic Billing Scenarios:
Traditional dynamic billing strategies does not incorporate thephenomenon of preserving privacy of homes because they aremore concerned towards providing dynamic billing incentives.However, these approaches can raise serious concerns towardsprivacy of smart homes. Because nowadays, grid utility collectthese real-time values in order to predict future load alongwith management of demand response, but these real-timevalues can leak personal information of smart home users.For instance, these values can be fed to NILM techniquesthat can even predict appliance usage of a specific house ina specified time-slot. Therefore, it is important to integrateprivacy preservation mechanism in dynamic billing strategyto preserve privacy. In order to do so, we integrate notionof differential privacy with smart grid dynamic billing andpropose our DRDP mechanism in this article.
2) Problem Questions:
We further divided the problemdefinition of DRDP mechanism into three critical points men-tioned as follows: • How to incentivize cooperating users that are not respon-sible to cause peak factor in a particular time-slot? (cf.Section IV-B1c) • How to preserve privacy of smart meters users alongside giving them advantages of dynamic billing? (cf.Section IV-B1a) • How to integrate the notion of differential privacy withusage based dynamic billing to provide smart homes witha billing strategy they can trust without worrying aboutprivacy leakage? (cf. Section IV-B1)
C. System Model
The proposed system model of our DRDP mechanismcomprises of two major entities e.g., smart homes and gridutility. Smart homes are entities which uses energy sent bysmart generation plants to carry out daily operations. Gridutility is the entity responsible to receive protected live updatesfrom smart homes to carry vary out billing processes. Gridutility is also responsible for storing data from all smart metersin their databased for future statistical tasks, such as DSM,load forecasting, etc.A detailed system model is given in Fig. 1, where everysmart home is linked with grid utility for real-time billingand monitoring purpose. Smart homes are equipped withsmart meters which record and accumulate their real-timeusage as instantaneous values ( I V ). After every 10 minutes,smart meters compute a differentially private noise ( M n ) fromLaplace distribution and add the generated noise to ( I V ) to getthe protected metering value ( P v ). Afterwards, smart metersreport this protected metering value ( P v ) to grid utility forbilling and other statistical operations. Grid utility has twomajor operation, first is calculation of dynamic billing andother is carrying out statistical analysis.In the first operation, grid utility provides fair dynamic pricingto all smart homes depending upon their usage. Grid utilityfirst works over adjustment of these reported values to findthe appropriate billing value. Afterwards, it gathers all real-time values ( P v ) in an specified area and calculates their sumto determine whether the usage for the specific area is largerthan peak value or not, in case if its larger than peak value,then it notifies all smart homes that peak-factor is in place,and warns smart homes to use minimum amount of energy.Moreover, it do also keeps track that whether a specific houseis responsible to cause peak-factor or not. In case, if a specificsmart home is consuming larger than average electricity value,then that smart home will be charged peak-price. Otherwise,the participating houses are only charged the normal price. Adetailed demonstration about this price calculation is given inDRDP algorithm demonstration section.In the second operation, grid utility carries out all statisticaltasks along with managing load for all areas. Grid utilitymanages collected real-time usage data to formulize loadcurves for future load. Similarly, it also manages grid powerstations and provides required instructions regarding differentbilling scenarios to according to each area. D. Adversary Model
Adversary in our model can be an intruder that is tryingto understand the real-time usage pattern of smart homesby analysing their reported readings ( P v ). To demonstrate itfurther, adversaries are actually interested to find out moreabout the lifestyle of smart homes users. Adversaries canbe of two types: 1) Harmless adversaries, which are justinterested to know usage patterns to carry out harmless tasksuch as targeted advertising after getting information about anydamaged appliance in smart home. 2) Harmful adversaries,which can cause serious threats to lives of smart home usersand can analyse the valuations to carry out unethical tasks TABLE II K EY N OTATIONS , D
ESCRIPTION , AND T HEIR V ALUE
Notation Description Value F n Function of Noise -
ABS
Absolute - I v Instantaneous MeteringValue - D f Difference Value - B R Billing Reading - S c Noise Scale - N r Number of Readings - I B Instantaneous Bill - M n Metering Noise - G SN Grid Side Noise - t s Time Slot - P v Protected Value - µ Mean Value - P F Peak Factor Value 12000Wh P P Price at Peak Time ¢25 U P Unit Price ¢10 N No. of Smart Meters 10 ∆ f Function Sensitivity 1 ε Epsilon (Privacy Parameter)
Multiple such as burglary, and theft, etc.We further divide the adversarial attacks in our DRDP mech-anism into two categories: 1) external attack from adversaries,in which adversarial attacker attacks the link of communicationbetween smart home and smart grid utility in order to find outdetailed usage information about homes in a specific region. 2)Internal adversarial attack, in which some internal grid entityact as an adversarial body and misuses the collected data fromgrid utility. Since grid utility databases have large amount ofdata from all local regions, they can pose large harm in casethey act as adversary. Furthermore, in this scenario, we assumethat adversary is curios-but-honest, as it will not modify, norwill alter or delete the received smart homes readings.IV. DRDP M
ECHANISM AND I TS COREF
UNCTIONALITIES
A. Preliminaries of DRDP1) Differential Privacy:
The notion of noise addition basedprivacy preservation also known as differential privacy wasfirst introduced by Cynthia Dwork in 2006 as a mean to protectdatabase privacy [18], [19]. Differential privacy works on theconcept of addition of i.i.d noise to obstruct malicious adver-saries from recovering private data from sensitive datasets [20].The notion was first used in statistical databased, but later onresearches identified that it also provides fruitful results whenit is used on real-time data [21]. In this article we use i.i.d noisegenerated from Laplace mechanism of differential privacy inorder to preserve privacy of smart metering real-time data. Theformal definitions of differential privacy are as follows:
Definition 1 (Adjacent Datasets)
In a given database D n consisting of n-dimensions, a queryfunction Q will provide ε -differential privacy P d if ∀ I , I ∈ D n vary by only a single element and all elements of R ∈ range ( Q ) [22]. Where R is the output value, D is Algorithm 1
Algorithm for Differentially Private DynamicPricing for Demand Response Enhancement while DifferentialDynamicPricing do Input ← F n , N, P F , U p , P p , I V , ε, µ, ∆ f Output ← P v , B R , I B , D f FUNCTION → Differential˙Reporting ( N, I v , ε , µ, ∆ f ) for (each i in N ) do Initialize
Mean ( µ ), epsilon ( ε ), sensitivity ∆ f Calculate
Scale S c1 = ∆ f ε Calculate
Noise =
Lap ( I v i , µ, S c1 ) Set
Meter Noise = M n = ABS [ Lap ( I v i , µ, S c1 )] Set
Protected Value = P v = I V i + M n end for return P v FUNCTION → Differential˙Noise˙Adjustment ( N, P v , ε , µ, ∆ f ) for (each i in N ) do Initialize
Mean ( µ ), epsilon ( ε ), sensitivity ∆ f Calculate
Scale S c2 = ∆ f ε Calculate
Noise =
Lap ( P v i , µ, S c2 ) Set
Grid Side Noise = G SN = ABS [ Lap ( P v i , µ, S c2 )] Set
Bill Reading = B R i = P v i − G SN end for return B R FUNCTION → Dynamic˙Billing ( N, P F , B R , U p , P p ) for (each i in N ) do Set
Sum = P B R i end for if Sum ≥ P F then Set
Avg = P F /N for (each j in N ) do if I V j ≥ Avg then I B j = B R j ∗ P P D f = B R j − Avg else I B j = B R j ∗ U P D f = Avg − B R j end if end for else for (each K in N ) do I B k = I B k ∗ U p end for end if return I B , D f end while designated database, and Q is the requested function of querythat satisfies ε -differential privacy. P d [ Q ( I ) ∈ R ] ≤ e ε × P d [ Q ( I ) ∈ R ] (1)In the above, range ( Q ) is the possible range for output valueof function Q . Correspondingly, the term ε is the privacyparameter used to determine the amount of noise whichis directly linked with the privacy level [23], [24]. Fromperspective of real-time data obfuscation of smart grid, weuse the concept of point-wise differential privacy, which wasfirst introduced by Eibl et al. in [25]. Definition 2 (Point-wise Sensitivity)
Every real-time value can be counted as an independent entity,and this value can be obfuscated individually on the basis ofits current attributes without linking it with its neighbouring value. The formal equation for point-wise sensitivity is givenas follows: ∆ P W ( f ) = max t s ,i ,i | f t S ( i ) − f t s ( i ) | = max i,t S | X i,t s | (2)In our DRDP mechanism, data obfuscation is carried outusing the concept of point-wise obfuscation mentioned inEq. 2. Furthermore, the sensitivity parameter ( ε ) controls thelevel of noise for any particular smart meter in a specific timeslot ( t s ). The value of ε can be varied according to the need,however, this value cannot be taken negative. For interestedaudience, a more detailed discussion regarding differentialprivacy can be found in [26].
2) Demand Response & Dynamic Billing:
DSM can for-mally be defined as a method to alter smart home usageprofiles in order to match them with the energy supplies [27].Similarly, DSM techniques are also being used to reduce op-erational cost, overcoming black outs, and to reduce emissionsof CO [28]. Among all DSM mechanisms, DR managementis considered to be the most prominent one to maintain abalance between load and supply curve. DR programs aredesigned and deployed in modern smart grids to enhanceparticipation of smart homes in load balancing. Many typesof DR mechanisms have been discussed in literature such ascontrol based, offer based, and decision variable based [5].Among these mechanisms, offer based mechanisms got thesignificant amount of attention because they directly incentiveusers and users can directly see their participation [29].In offer based DR models, motivation is developed amongsmart homes to use minimal amount of energy in the giventime slot so that grid utility can balance the load curve andcan predict the load in the most proficient manner [30]. In thisarticle, we use a subcategory of offer based DR mechanism inwhich we provide incentives to participating users on the basisof the factor that whether they are contributing in causing peakfactor or not. B. Functioning of DRDP Mechanism
We divide the discussion about functioning into two parts,one part discusses the algorithm and the other part discussesthe theoretical analysis of the proposed work.
1) DRDP Algorithm:
The proposed DRDP algorithm canfurther be split into three parts, first two discuss about dif-ferential privacy, while the third talk about incentivizing ofsmart homes. In this section, we discuss all three parts fromtechnical perspective. a) DRDP Private Reporting:
In order to protect theinstantaneous values ( I V ) of smart meters, we use the phe-nomenon of differentially private noise addition using Laplacedifferential privacy mechanism. The pseudo-code for noiseaddition is given in first Function of Algortihm 1. The noisevia Laplace is calculated as follows: Lap ( I V , µ, S c1 ) = f ( I V , µ, S c1 ) = 12 S c1 e | IV − µ | Sc1 (3)Similarly, the above equation can further be broken down fordetailed understanding by substituting the value of ( S c1 = ∆ f ε ) , the new equation will be: f (cid:16) I V ; µ, ∆ f ε (cid:17) = ∆ f ε .e − | I V − µ | ∆ f ε ! (4)The calculated noise is then added into the instantaneous valueof smart meters to produce a noisy output as follows: N X i =0 ( P v i = I V i + ABS [ Lap ( I v i , µ, S c1 )]) (5)Finally, this protected noisy value is then sent to smart gridutility for billing, storage, and future statistical evaluation.Grid utility first works over adjustment of noisy values forbilling calculation and then carry out various statistical anal-ysis over these reading such as carrying out load forecasting,etc. It is important to highlight that the protected noisyinstantaneous values does not have any significant effect onbilling or load forecasting as far as the ε value is maintainedaccordingly. b) Differential Noise Adjustment: The second functionin DRDP mechanism is differential noise adjustment, via thisfunction, grid utility generate a random i.i.d noise at its endand reduces this noise value from the reported reading in orderto match the accurate value for billing. Firstly, the noise isgenerated at the grid utility end by using similar Laplace noisemechanism [31]. Usually the epsilon value is the same as thatof smart meter end, but it can be varied if required. The formaldistribution used at grid utility end is as follows:
Lap ( P v , µ, S c2 ) = f ( P v , µ, S c2 ) = 12 S c2 e | Pv − µ | Sc2 (6)The generated noise is then reduced from protected value togenerate the final reading value for billing and future analysis.The equation for this process is as follows: N X i =0 ( B R i = P v i − ABS [ Lap ( P v i , µ, S c2 )]) (7)It is important to mention that it is not compulsory that thenewly generated value ( B R ) will always match the originalvalue ( I V ), because there is always a possibility that the newnoise value could be pretty small or pretty large as compared tothe original noise value generated at meters’ end. In majorityof adjustments, both the original and new generated valuesare pretty similar. But there will always remain a sense ofambiguity and uncertainty in output values even after removalof noise. This introduction of ambiguity is the actual require-ment of any differential privacy mechanism, that an adversarywill not be able to predict with confidence regarding presenceor absence of any individual. In our scenario, if an adversaryeven gets the corrected values ( B R ) from grid utility, even thenthese values will be of no use to NILM mechanisms. As theseNILM mechanisms will not be able to predict with confidenceregarding presence or absence of any specific appliance insmart home because of the noise ambiguity factor. On theother hand, this noise adjustment does not have much effectof billing values and results have shown that a very minimal level of error is found in billing, which can be ignored becauseof being very small. c) Incentivizing Participating Homes: Conventionalmechanisms of dynamic pricing charge all users the same priceirrespective of their usage, however, in our proposed DRDPmechanism we only charge homes who are causing the peakfactor in that specific region. Because of this, cooperative userswill have a motivation to take part in DSM programs, whicheventually will have beneficial effect on load curve. The thirdfunction of our proposed DRDP algorithm (Algorithm. 1)determines and calculate dynamic bill for each smart home.Firstly, adjusted readings values ( B R ) of all smart homes iscollected and sum of all these values is computed via gridutility ( sum = P N B R ) . Afterwards, grid utility derives thepeak factor value ( P F ) for that specific time-slot according tothe load curve given by grid utility. Once P F is determined,utility compares the summation value ( sum B R ) with peakvalue ( P F ) to determine that whether the instantaneous sumof all smart homes exceeds peak or not ( sum B R ≥ P F ). Incase if instantaneous sum is greater than the selected peakfactor, then the smart meter homes are given a notificationthat peak factor is in place and energy usage is being chargedaccording to the peak prices. Along with the peak factorcomparison, grid utility also calculates instantaneous averagevalue according to the peak factor and number of smart homesvia ( avg = P F n . The instantaneous value of each smart home B R is then compared with this calculated average and in caseif the smart home is using energy higher than average, thenthey are charged for peak prices, for example if N ∗ are thesmart homes using energy larger than average, then the billingwill be as follows: N ∗ X i =1 I B i = N ∗ X i =1 ( B R i × P F P ) (8)Contrary to this, if some smart home is participating in DSMand is using less energy, then they are charged off-peak price( U OP ) as follows: N p X i =1 I B i = N p X i =1 ( B R i × U OP ) (9)We further add the phenomenon of communicating smarthomes regarding the energy difference with respect to averagevalue. For example, if a meter is only using 50W more thanpeak value, which they can reduce, or a smart home is just10W short from reaching peak value and they do not wantto get into peak zone. In order to do so, we calculate thedifference between their instantaneous reading ( B R ) and theaverage ( Avg ) via ( D f = B R – Avg ) for peak users and ( D f = Avg – B R ) for non-peak users. This calculated difference valuewill then be transmitted to respective smart home to notifythem about their usage.
2) Theoretical Analysis:
We divide theoretical analysis intothree major parts; firstly, we provide extensive theoreticalanalysis for differential privacy guarantees. Then, we analysecooperative state in our DRDP model, and afterwards, weprovide complexity analysis for the proposed algorithm. a) Differential Privacy Analysis:
In order to demonstratethat our proposed noise addition mechanism follows differ-ential privacy guarantee, we carry out extensive theoreticalevaluation. We evaluate both of the proposed functions, firstwe proved that noise addition at smart metering side satisfiesdifferential privacy guarantee. Afterwards, we prove that ournoise adjustment mechanism at grid utility end also satisfieddifferential privacy guarantee. The detailed evaluation is givenas follows:
Theorem
1: Differentially private metering reporting func-tion of our proposed DRDP mechanism satisfies ε -differentialprivacy guarantee. Proof:
Let us consider M n & M ′ n ∈ N | X | in a way that || M n − M ′ n || ≤ . The arbitrary string length up to ‘ i ′ for M n & M ′ n will be M = { N , N , .., N i } . Thus, given that both M n & M ′ n can further be linked with Laplace distribution viaprobability density function as p M n & p M ′ n respectively. Thesetwo probability functions can be compared at given arbitrarystring (according to Laplace theorem in [32]) as follows: p M n [ M = { N , N , .., N i } ] p M ′ n [ M = { N , N , .., N i } ] = (10) k Y j =1 exp (cid:16) − ε | F n ( M n ) j − N j | ∆ f (cid:17) exp (cid:16) − ε | F n ( M ′ n ) j − N j | ∆ f (cid:17) (11) = k Y j =1 exp (cid:18) ε ( | F n ( M ′ n ) j − N j |−| F n ( M n ) j − N j | )∆ f (cid:19) (12) ≤ k Y j =1 exp (cid:18) ε ( | F n ( M n ) j − | F n ( M ′ n ) j | )∆ f (cid:19) (13) = exp (cid:18) ε ( || F n ( M n ) − | F n ( M ′ n ) || )∆ f (cid:19) (14) ≤ exp( ε ) (15)Thus, the above statements prove that differentially privatereporting of our DRDP satisfies ε –differential privacy. Since,in real-time reporting, we are taking noise values to accumu-late in I v , so, the given differential privacy function followingpositive side of noise symmetry. Theorem
2: Differential noise adjustment function of ourproposed DRDP mechanism satisfies ε -differential privacyguarantee. Proof:
Let us consider G SN & G ′ SN ∈ N | X | in a waythat || G SN − G ′ SN || ≤ . The arbitrary string length up to ‘ i ′ for G SN & G ′ SN will be G S = { G , G , .., G i } . Thus, giventhat both G SN & G ′ SN can further be linked with Laplacedistribution via probability density function as p G SN & p G ′ SN respectively. These two probability functions can be comparedat given arbitrary string (according to Laplace theorem in [32])as follows: p G SN [ G S = { G , G , .., G i } ] p G ′ SN [ G S = { G , G , .., G i } ] = (16) k Y j =1 exp (cid:16) − ε | F n ( G SN ) j − N j | ∆ f (cid:17) exp (cid:16) − ε | F n ( G ′ SN ) j − N j | ∆ f (cid:17) (17) = k Y j =1 exp (cid:18) ε ( | F n ( G ′ SN ) j − N j |−| F n ( G SN ) j − N j | )∆ f (cid:19) (18) ≤ k Y j =1 exp (cid:18) ε ( | F n ( G SN ) j − | F n ( G ′ SN ) j | )∆ f (cid:19) (19) = exp (cid:18) ε ( || F n ( G SN ) − | F n ( G ′ SN ) || )∆ f (cid:19) (20) ≤ exp( ε ) (21)Thus, the above statements prove that differential noiseadjustment function of our DRDP satisfies ε -differentialprivacy. Since, in noise adjustment, we are taking removingnoise values in order to match the correct values of I v as muchas possible. So, the given noise adjusting differential privacyfunction following negative side of noise symmetry.Both Theorems 1 & 2 can be combined to prove that bothside of symmetries of Laplace distribution are followed in ourDRDP model. Theorem
3: Differential noise adjustment function of ourproposed DRDP mechanism satisfies ε -differential privacyguarantee. Proof:
In our proposed DRDP algorithm, Laplacedistribution of differential privacy is applied in a sequentialstep-wise manner via ε & ε privacy budgets. Thus, by fol-lowing composition theorem of differential privacy accordingto Lemma 1, if we perform sequential perturbation on samesmart metering data by using ε & ε , then, they can be accu-mulated via summation to prove differential privacy guarantee(e.g., P j ε j − dp ). Therefore, our proposed differential noiseaddition (using ε ) and differential noise adjustment (using ε ) of DRDP can be written as ( ε + ε )-differential privacyto prove guarantee via composition theorem. b) Cooperative State Analysis: Considering the systemmodel and functioning given in previous sections, it can bevisualized that at the time of peak-factor in place, two typesof behaviours of smart meter nodes can be seen. Either theyare in cooperative state (e.g., using less than average) orin non-cooperative state (using more than average). Basedon these conditions, we devise two states of systems namedas cooperative and non-cooperative. To clear it further, wedeveloped the notion that when at least half of the smart meterswill be in a cooperative state and using less than averagevalue, then the complete network will be in a cooperativestate. Contrary to this, if less than half of the smart meteringnodes are using more than average, then the system willbe in a non-cooperative state. These states can be used todetermine the future response of the system based upon itscurrent sates. In order to elaborate it further, we carry outextensive theoretical analysis of these states using probabilitytheory concepts, which is explained further.
Theorem
4: The probability of system being in cooperativestate is: (22) P cs = N X q = ⌈ N ⌉ (cid:18) Nq (cid:19) (cid:16) P ( m ) LU (cid:17) q (cid:16) P ( m ) HU (cid:17) N − q In the above equations, P LU and P HU are cooperativeand non-cooperative user probabilities respectively which aredemonstrated in detail below. Proof:
Considering the factor that a smart meter can either bein a cooperative or a non-cooperative sate we determine thestate probability vectors as follows: P LU = { P L (1) , P L (2) , P L (3) , ..., P L ( N ) } P HU = { P H (1) , P H (2) , P H (3) , ..., P H ( N ) } Consider S m be the binomial random variable for smartmeters in a cooperative state, then P( S m = q ) will be theprobability that q number of nodes in cooperative state duringpeak-hours, which can be written as follows: (23) P { S m = q } = (cid:18) Nq (cid:19) (cid:16) P ( m ) LU (cid:17) q (cid:16) − P ( m ) LU (cid:17) N − q The system will remain in non-cooperative state unless ⌈ N ⌉ smart meters enters in cooperative state, so, the probability ofbeing in non-cooperative state can be calculated from Eq. 23as follows: (24) P NC = ⌊ N ⌋ X q =0 (cid:18) Nq (cid:19) (cid:16) P ( m ) LU (cid:17) q (cid:16) − P ( m ) LU (cid:17) N − q Complying with the probability condition that ( P CS + P NC = 1) , the above equation can be written as: (25) P CS = 1 − P NC = 1 − ⌊ N ⌋ X q =0 (cid:18) Nq (cid:19) (cid:16) P ( m ) LU (cid:17) q (cid:16) − P ( m ) LU (cid:17) N − q According to the probability vectors of P LU and P HU , in-dividual values of each vector can be compared accordingto the probability condition of summation equal to 1 (e.g., P L (1) + P H (1) = 1 , which can further generalized for abovesummation as P ( m ) LU + P ( m ) HU = 1 . So, Eqn. 25, will become:(26) P CS = 1 − ⌊ N ⌋ X q =0 (cid:18) Nq (cid:19) (cid:16) P ( m ) LU (cid:17) q (cid:16) P ( m ) HU (cid:17) N − q The above equation provides the probability of system beingin cooperative state, which means that at least ⌈ N ⌉ nodes arein cooperative state. So, the Eqn. 26 can be modified to provethe theorem as follows: (27) P cs = N X q = ⌈ N ⌉ (cid:18) Nq (cid:19) (cid:16) P ( m ) LU (cid:17) q (cid:16) P ( m ) HU (cid:17) N − q Moreover, Eq. 27 can be used to determine the expectedvalue of smart meters, which is used to determine the expectednumber of smart meter nodes in cooperative state at differentprobability values. So, the equation for expectation can bederived from Eq. 27 as: (28) E [ P CS ] = N X q = ⌈ N ⌉ q. (cid:18) Nq (cid:19) (cid:16) P ( m ) LU (cid:17) q (cid:16) P ( m ) HU (cid:17) N − q c) Complexity Analysis: The proposed DRDP Algorithm for real-time private reportingand smart dynamic billing is computationally efficient as it uti-lizes minimal required amount of operations for its execution.
Theorem
5: The computational complexity of our proposedDRDP Algorithm has an upper bound of O ( N ) .Proof: The propose DRDP algorithm can be divided intothree sub-algorithms, which are executed sequentially and arenamed as differential reporting, differential noise adjustment,and differential dynamic billing. The first sub-algorithm inDRDP is differential reporting, which calculates and addsdifferential noise for each smart meter and report this value togrid utility. This part mainly comprises of ‘for’ loop, which isiterated for ‘N’ number of times. It is important to note that‘for’ loops in out algorithm have computational complexity of O [3( N + 1)] . This complexity value of O [3( N + 1)] is furtherrounded to O ( N ) for simplicity. The statements from Line5 - Line 9 comprise of initialization, noise computation, andaddition, these statements are considered as computationallyconstant because they are executed only once in the ‘for’ loop.Therefore, they have computational complexity of O (1) . So,this part of DRDP has an upper bound of O ( N ) computationalcomplexity.Similarly, the second ‘differential noise adjustment’ part ofDRDP algorithm has similar computational complexity O ( N ) because of similar statements present in it, with just a slightdifference of reduction of noise value on Line 17. However,this difference does not make any variance in computationalcomplexity, and overall upper bound computational com-plexity of noise adjustment mechanism is also O ( N ) . It isimportant to note that the exponential function used in noisedetermination from Laplace distribution on Line 7 & Line 13are only executed once. These are not continuous exponentialand only execute after getting a fixed set of inputs, therefore,the complexity of these exponential statements is consideredas constant O (1) .The third dynamic billing part of DRDP further comprisesof two steps, first is computes sum and the compare it withreading for billing price. In the first part, for loop is executedfor ‘N’ number of times to calculate sum, so, it has compu-tational complexity of O ( N ) . Afterwards, an ‘If’ condition isused to determine the occurrence of peak factor, which is acontact statement and have complexity of O (1) . In case ofboth positive and negative ‘if’, a ‘for’ loop is executed for an‘N’ number of times which have computational complexity of O ( N ) . All these statements can further be combined to statethat the upper bound of computational complexity of dynamicbilling function is O ( N ) .Keeping in view, the upper bounds of all three sub-algorithms,it can be stated that the upper bound computational complexityof our proposed DRDP mechanism is O ( N ) . Fig. 2: Performance Evaluation of Noisy Reporting Functionof DRDP Mechanism. The graph shows absolute private valuesreported to smart grid utility from smart meter after additionof differentially private noise at different epsilons ( ε ) values.Fig. 3: Analysis of Mean Absolute Error (MAE) Added in eachMeter Reading with Respect to Privacy Budget ( ε ). The valuesof MAE are absolute error values and are not in percentage.V. P ERFORMANCE E VALUATION OF
DRDPIn this section, we provide the evaluation results of ourproposed DRDP algorithm along the comparison with similarstate-of-the-art works. To evaluate our DRDP mechanism, wetook the dataset of [33], and further modified it with respectiveto grid utility, and smart homes scenario. Furthermore, forcomparison, we gained the UDP idea from work carried outin [3]. To perform experimental evaluation, we made use of thelibrary NumPy from Python 3.0, and performed experimentsover smart meter transmitted data having an interval of 10minutes between each reading. The simulation parameters usedin our experiment are provided in Table. II. Fig. 4: Accumulated Sum of All Participating Homes afterNoise Adjustment via DRDP.Fig. 5: Accumulated Billing Sum for 10 Homes Using afterusing Incentivized Dynamic Billing of DRDP on AdjustedNoise Values Reported at Different Privacy Budgets.We further divide the experimental evaluation into three parts,first we analyse DRDP strategy from perspective of differentialprivacy noise addition and adjustment, afterwards, we analysethe dynamic billing, and finally we evaluate the cooperativesmart home analysis.
A. Private Grid Reporting and Noise Adjustment
The graphs presented in Fig. 2, 3, and 4 demonstrate thenoise reporting and adjustment scenario. Firstly in Fig. 2, twographs are shown, the first graph demonstrate the real-timereadings reported from smart meter to grid utility. The graphis built using 3 days data of a smart home readings at differentvalues of privacy parameters. The second graph in the figure isthe zoomed version of first graph, which is zoomed in orderto visualize the changes due to added noise. From both ofthe figures, it can be seen that the addition of noise distortthe original values for privacy protection. Especially, whenthe value of ε is small, large distortion can be seen at thevalues, which means more privacy is preserved. Fig. 3 can bevisualized in order to find out the error rate at each ε value.Mean absolute error (MAE) in our DRDP is calculated bytaking the sum of absolute difference between noisy valuesand original values for a smart home through out the reported Fig. 6: Billing Graphs for a Randomly Picked Smart Home inorder to Visualize the incentive given by DRDP as Comparedto UDP Pricing [3].Fig. 7: Evaluation of Deviation Notification Function of DRDPfrom Each Bill reading for a Smart Home.time, and then this accumulated difference is divided by thetotal readings involved in the experiment (e.g., 3 days in ourexperiments) ( M AE = P Nrn =1 | P v − I v | N r ). From Fig. 3, it canbe seen that the value of MAE is highest at the time of ε =0.01, and it tend to reduce with the increasing value of ε . Itis important to mention that lower values does not mean thatprivacy is not preserved, as the lower values do also preserveprivacy of smart meters from NILM strategies to a greaterextant as NILM strategies cannot predict with confidence dueto added noise.Moving further to noise adjustment, the first graph demon-strating the effect on accumulated meter reading can be seenin Fig.4. The given graph demonstrates the summation of allhomes for 3 days after noise adjustment. It can be seen fromthe graph, that even after summation of values from all 10homes, very minimal difference can be seen with respect tooriginal value ‘without proposed DRDP’. Which means thatthe adjusted values are pretty close to the original values,which directly means that the error in the billing value willbe very minimal which can be neglected by considering itas a trade-off of preserving privacy. These adjusted valuesare then fed to billing function for bill calculation, which isdemonstrated in the next part of performance evaluation. Fig. 8: Expectation of Smart Homes for Cooperative StateAnalysis at Different Probability Values.
B. Incentivized Billing Evaluation
Since, billing is another major aspect of our contribution,so, we demonstrate this functionality by showing as muchas possible experimental results in Fig, 5, 6, 7. The majorconcern while calculating billing from noisy values was thatit will have huge error, and this will not be able to matchwith original values. However, we overcome this concernby proposing noise adjustment function, and we evaluate itsusefulness at different ε values and showed it in Fig. 5. Theaccumulated bill shown in the figure is calculated by accu-mulating billing values of all smart homes within a timespanof three days. In the given figure, first bar in red textureshows the proposed dynamic billing strategy but without noiseaddition. However, the remaining bars show the accumulatedreading of a smart home by using noisy values at differentprivacy budgets. From the results, it can be visualized thatthere is very minimal difference in bills of all smart homes.Even at ε = 0.01, when the value of noise is pretty high atthe time of reporting, even then the accumulated bill of allsmart homes have very low or no variance with respect tooriginal bill. These results demonstrate the effectiveness of ournoise adjustment function, that from surface level one can takeperception that the noisy value might cause bill billing error,but this did not happened. Contrary to this, in the long-run theoverall billing difference is negligible. So, we firmly believethat our proposed mechanism can be implemented in real-timesmart meters to protect their privacy along side providing themusage based dynamic billing.Furthermore, in Fig. 6, the separated billing graph for onesmart home can be visualized. In the given graph, it can bevisualized that our proposed DRDP mechanism only chargesthe user when they are causing the peak factor and it does notcharges when the specific home is not causing peak factor.E.g., from 07:30PM to 09:00PM, it can be seen that the homewas not responsible for causing the peak factor, however, theUDP strategy charged the smart home with peak price. Sameresult can be seen at almost everyday in the time-slot from07:30PM to 09:00PM, as the home is generally cooperatingin these slots because the peak factor has occurred. Therefore,according to DRDP strategy, it is being charged low pricebecause of its cooperation, however, in UDP mechanism, it isbeing charged according to the same tariff as that of other smart homes. The third graph in Fig. 7 shows the outputof deviation function that we added in our enhanced pricingmodel. This function calculates the difference of a smart meterfrom the average value and report the difference to smart meteruser in order for him to take adequate action. E.g., in case ifpeak factor is in place and a smart home is just using few wattsless than peak value, then it is notified that you are ‘X’ amountshort from reaching average value. This notification is like aninitial alert message to smart home user, which afterward tryto control its usage a bit further in order to not fall above peakfactor. C. Cooperative State Evaluation
From the perspective of cooperative state analysis, weprovide experiments results in Fig. 8. In the given figure,expected value of number of smart meters has been shown atdifferent probability values. For example, in case of 12 smartmeters, the expectation is minimum at p = 0 . , however, thesame values reached approximately the maximum limit at p = 0 . . The same trend can be visualized for other numberof smart meters as well, which can be used to conclude thathigher the probability value, higher will be the expectednumber of smart meters in cooperative state. Hence, after careful analysis of the experimental results pro-vided in experimental graphs, it can easily be determined thatDRDP mechanism efficiently provide smart metering privacyalong with providing benefit to cooperative users in dynamicbilling scenario.
VI. C
ONCLUSION
In this paper, we enhance traditional dynamic billing mech-anism for smart homes by providing an incentivising dy-namic pricing mechanism for cooperative users. Furthermore,we provide a differentially private reporting mechanism forsmart meters to protect their privacy. Collectively, a D emand R esponse enhancing D ifferential P ricing (DRDP) mechanismhave been proposed to be incorporated into smart meters andgrid utility for demand side management. A detailed theoret-ical analysis have been carried out for our proposed DRDPmechanism. Similarly, extensive performance evaluation atdifferent privacy parameters have been carried out as well.The provided analysis and performance evaluation show thatour proposed DRDP mechanism outperforms traditional andstate-of-the-art works in dynamic pricing and private smartmetering. R EFERENCES[1] P. Kumar, Y. Lin, G. Bai, A. Paverd, J. S. Dong, and A. Martin, “Smartgrid metering networks: A survey on security, privacy and open researchissues,”
IEEE Communications Surveys & Tutorials , vol. 21, no. 3, pp.2886–2927, 2019.[2] Z. Ma, H. Zhong, Q. Xia, and C. Kang, “A block-of-use electricity retailpricing approach based on the customer load profile,”
IEEE Transactionson Smart Grid , vol. 11, no. 2, pp. 1500–1509, 2020.[3] P. Gope and B. Sikdar, “An efficient data aggregation scheme forprivacy-friendly dynamic pricing-based billing and demand-responsemanagement in smart grids,”
IEEE Internet of Things Journal , vol. 5,no. 4, pp. 3126–3135, 2018. [4] M. Majidi and K. Zare, “Integration of smart energy hubs in distributionnetworks under uncertainties and demand response concept,” IEEETransactions on Power Systems , vol. 34, no. 1, pp. 566–574, 2019.[5] J. S. Vardakas, N. Zorba, and C. V. Verikoukis, “A survey on demandresponse programs in smart grids: Pricing methods and optimizationalgorithms,”
IEEE Communications Surveys & Tutorials , vol. 17, no. 1,pp. 152–178, 2015.[6] K. Chen, Y. Zhang, Q. Wang, J. Hu, H. Fan, and J. He, “Scale-and context-aware convolutional non-intrusive load monitoring,”
IEEETransactions on Power Systems , vol. 35, no. 3, pp. 2362–2373, 2020.[7] M. U. Hassan, M. H. Rehmani, R. Kotagiri, J. Zhang, and J. Chen, “Dif-ferential privacy for renewable energy resources based smart metering,”
Journal of Parallel and Distributed Computing , vol. 131, pp. 69–80,2019.[8] X. Liang, X. Li, R. Lu, X. Lin, and X. Shen, “Udp: Usage-based dynamicpricing with privacy preservation for smart grid,”
IEEE Transactions onSmart Grid , vol. 4, no. 1, pp. 141–150, 2013.[9] J. Ferdous, M. P. Mollah, M. A. Razzaque, M. M. Hassan, A. Alamri,G. Fortino, and M. Zhou, “Optimal dynamic pricing for trading-off userutility and operator profit in smart grid,”
IEEE Transactions on Systems,Man, and Cybernetics: Systems , 2017.[10] Z. Almahmoud, J. Crandall, K. Elbassioni, T. T. Nguyen, andM. Roozbehani, “Dynamic pricing in smart grids under thresholdingpolicies,”
IEEE Transactions on Smart Grid , vol. 10, no. 3, pp. 3415–3429, 2018.[11] J. Won, C. Y. Ma, D. K. Yau, and N. S. Rao, “Privacy-assured aggrega-tion protocol for smart metering: A proactive fault-tolerant approach,”
IEEE/ACM Transactions on Networking (TON) , vol. 24, no. 3, pp. 1661–1674, 2016.[12] J. Ni, K. Zhang, K. Alharbi, X. Lin, N. Zhang, and X. S. Shen,“Differentially private smart metering with fault tolerance and range-based filtering,”
IEEE Transactions on Smart Grid , vol. 8, no. 5, pp.2483–2493, 2017.[13] J. Zhao, J. Liu, Z. Qin, and K. Ren, “Privacy protection schemebased on remote anonymous attestation for trusted smart meters,”
IEEETransactions on Smart Grid , vol. 9, no. 4, pp. 3313–3320, 2016.[14] P. Zeng, H. Li, H. He, and S. Li, “Dynamic energy management of amicrogrid using approximate dynamic programming and deep recurrentneural network learning,”
IEEE Transactions on Smart Grid , vol. 10,no. 4, pp. 4435–4445, 2019.[15] A. Mohammadali and M. S. Haghighi, “A privacy-preserving homomor-phic scheme with multiple dimensions and fault tolerance for meteringdata aggregation in smart grid,”
IEEE Transactions on Smart Grid , pp.1–1, 2021.[16] W. Kong, Z. Y. Dong, J. Ma, D. J. Hill, J. Zhao, and F. Luo, “Anextensible approach for non-intrusive load disaggregation with smartmeter data,”
IEEE Transactions on Smart Grid , vol. 9, no. 4, pp. 3362–3372, July 2018.[17] M. U. Hassan, M. H. Rehmani, and J. Chen, “Differentially privatedynamic pricing for efficient demand response in smart grid,” in
ICC2020 - 2020 IEEE International Conference on Communications (ICC) ,2020, pp. 1–6.[18] C. Dwork, “Differential privacy,” in
Automata, Languages and Program-ming . Berlin, Heidelberg: Springer Berlin Heidelberg, 2006, pp. 1–12.[19] F. Zhao, X. Ren, S. Yang, Q. Han, P. Zhao, and X. Yang, “Latent dirichletallocation model training with differential privacy,”
IEEE Transactionson Information Forensics and Security , vol. 16, pp. 1290–1305, 2021.[20] X. Cheng, P. Tang, S. Su, R. Chen, Z. Wu, and B. Zhu, “Multi-party high-dimensional data publishing under differential privacy,”
IEEETransactions on Knowledge and Data Engineering , vol. 32, no. 8, pp.1557–1571, 2020.[21] M. U. Hassan, M. H. Rehmani, and J. Chen, “Privacy preservation inblockchain based iot systems: Integration issues, prospects, challenges,and future research directions,”
Future Generation Computer Systems ,vol. 97, pp. 512–529, 2019.[22] L. Wang, D. Zhang, D. Yang, B. Y. Lim, X. Han, and X. Ma, “Sparsemobile crowdsensing with differential and distortion location privacy,”
IEEE Transactions on Information Forensics and Security , vol. 15, pp.2735–2749, 2020.[23] T. Zhu, D. Ye, W. Wang, W. Zhou, and P. Yu, “More than privacy:Applying differential privacy in key areas of artificial intelligence,”
IEEETransactions on Knowledge and Data Engineering, in Print , pp. 1–1,2020.[24] M. U. Hassan, M. H. Rehmani, and J. Chen, “Deal: Differentiallyprivate auction for blockchain-based microgrids energy trading,”
IEEETransactions on Services Computing , vol. 13, no. 2, pp. 263–275, 2020. [25] G. Eibl and D. Engel, “Differential privacy for real smart metering data,”
Computer Science-Research and Development , vol. 32, no. 1-2, pp. 173–182, 2017.[26] M. U. Hassan, M. H. Rehmani, and J. Chen, “Differential privacytechniques for cyber physical systems: A survey,”
IEEE CommunicationsSurveys Tutorials , vol. 22, no. 1, pp. 746–789, 2020.[27] M. Alizadeh, X. Li, Z. Wang, A. Scaglione, and R. Melton, “Demand-side management in the smart grid: Information processing for the powerswitch,”
IEEE Signal Processing Magazine , vol. 29, no. 5, pp. 55–67,2012.[28] Z. Fan, P. Kulkarni, S. Gormus, C. Efthymiou, G. Kalogridis,M. Sooriyabandara, Z. Zhu, S. Lambotharan, and W. H. Chin, “Smartgrid communications: Overview of research challenges, solutions, andstandardization activities,”
IEEE Communications Surveys & Tutorials ,vol. 15, no. 1, pp. 21–38, 2012.[29] W. Liu, D. Qi, and F. Wen, “Intraday residential demand responsescheme based on peer-to-peer energy trading,”
IEEE Transactions onIndustrial Informatics , vol. 16, no. 3, pp. 1823–1835, 2020.[30] Q. Shi, C.-F. Chen, A. Mammoli, and F. Li, “Estimating the profileof incentive-based demand response (ibdr) by integrating technicalmodels and social-behavioral factors,”
IEEE Transactions on SmartGrid , vol. 11, no. 1, pp. 171–183, 2019.[31] L. Ou, Z. Qin, S. Liao, Y. Hong, and X. Jia, “Releasing correlatedtrajectories: Towards high utility and optimal differential privacy,”
IEEETransactions on Dependable and Secure Computing , vol. 17, no. 5, pp.1109–1123, 2020.[32] C. Dwork, A. Roth et al. , “The algorithmic foundations of differentialprivacy.”
Foundations and Trends in Theoretical Computer Science ,vol. 9, no. 3-4, pp. 211–407, 2014.[33] M. Muratori, “Impact of uncoordinated plug-in electric vehicle chargingon residential power demand-supplementary data,” National RenewableEnergy Laboratory-Data (NREL-DATA), Golden, CO (United States),Tech. Rep., 2017.
Muneeb Ul Hassan received his Bachelor degreein Electrical Engineering from COMSATS Instituteof Information Technology, Wah Cantt, Pakistan,in 2017. He received Gold Medal in Bachelordegree for being topper of Electrical EngineeringDepartment. Currently, he is pursuing the Ph.D.degree from Swinburne University of Technology,Hawthorn VIC 3122, Australia. His research inter-ests include privacy preservation, blockchain, Inter-net of Things, decentralized IoT systems, securityand privacy issues, Ad-Hoc networks, cyber physicalsystems, smart grid, cognitive radio networks, and big data.
Mubashir Husain Rehmani (M’14-SM’15) re-ceived the B.Eng. degree in computer systems en-gineering from Mehran University of Engineeringand Technology, Jamshoro, Pakistan, in 2004, theM.S. degree from the University of Paris XI, Paris,France, in 2008, and the Ph.D. degree from theUniversity Pierre and Marie Curie, Paris, in 2011.He is currently working as Assistant Lecturer atMunster Technological University (MTU), Ireland.He worked at Telecommunications Software andSystems Group (TSSG), Waterford Institute of Tech-nology (WIT), Waterford, Ireland as Post-Doctoral researcher from Sep 2017to Oct 2018. He served for five years as an Assistant Professor at COMSATSInstitute of Information Technology, Wah Cantt., Pakistan.