Text from the interactive explanation
ElecPrivacy: Evaluating the Privacy Protection of Electricity Management Algorithms
Georgios Kalogridis, Rafael Cepeda, Stojan Z. Denic, Tim Lewis, Costas Efthymiou
Original paperElecPrivacy · IEEE Transactions on Smart Grid · 2011
A detailed smart-meter reading shows every appliance switching on and off — and with it, when you wake, cook, wash or go out. A Toshiba research team in Bristol asked whether a home battery could blur that picture, and whether residents could choose what to hide.
Switch something on
The eavesdropper's notebook
Try this: tap Kettle and watch the amber spike. Then switch to No battery and back. What does the meter reveal in each case? Start the story ↓
1 · Fingerprints
Every appliance leaves a fingerprint
The researchers recorded electricity the way a very detailed smart meter would: the average power over each 30-second slot , from ten wireless monitors attached to different groups of appliances in a studio apartment, for 13 days. Each appliance draws power in its own recognisable shape — the paper calls it a power load signature . Add every signature together and you get the home's total: the only thing a household meter sees.
Name that appliance Simulated traces
Each trace is one appliance on its own. Which one is it?
Pick an answer — the shape is the clue.
What a jump can reveal From the paper
The paper's simple example (its Table II): how strongly using each appliance suggests that someone is home.
The authors' privacy goal is "privacy of personal behavior": appliance events should be undetectable, because knowing that an appliance ran can imply whether a home is occupied and hint at routines such as sleeping times.
From the total to the jumps Simulated evening, 17:30–20:30
An eavesdropper doesn't need a monitor on each appliance. The paper assumes a deliberately strong adversary who knows every appliance's signature and can split the home's total into appliances by spotting edges — sudden jumps between one reading and the next, dp(t) = p(t) − p(t − Δt). Every jump, or lack of one, is treated as private information.
Would less frequent readings help? Same simulated day
The 30-second readings in this study are more detailed than the 15-minute readings a utility might collect. Drag the slider to average the same day over longer intervals and count the jumps that remain.
Coarser readings blur the details, yet the paper cites the argument that even 15-minute data may open "a window into the activities within homes". ElecPrivacy takes another route: it changes what the home sends to the meter in the first place.
2 · The battery trick
A battery as a shock absorber
ElecPrivacy puts a rechargeable battery (an electric-vehicle battery, for example) between the grid and the home, along with a "power router" that mixes battery power and grid power. When an appliance switches on, the battery can supply the extra power, so the meter barely moves; later, it recharges. The meter then sees
p′(t) = p(t) − pB(t) — the home's demand minus the battery's contribution.
Every real battery has two limits: how fast it can deliver or absorb power, and how much energy it holds. The authors' earlier "best-effort" algorithm simply keeps the meter reading constant for as long as those limits allow.
−PR ≤ pB ≤ PD0 ≤ energy used ≤ EC
A perfectly flat meter line forever would need an endlessly large battery, or knowledge of every future appliance use. Real algorithms must decide one reading at a time.
Predict, then run Single-event lab
Make a prediction, then press “Run the event”.
3 · The paper's new idea
Hide the kettle, not the washing machine
Hiding everything is costly, and some appliances aren't worth hiding: the fridge cycled on and off in the same way on weekdays and weekends, whatever the residents were doing. So the paper introduces a selective algorithm. One physical battery is split into logical energy quotas , one share for each appliance group you want to protect. By default the shares follow each group's share of energy use, leaving out the fridge and other background loads — together 59.92% of consumption.
Each share then runs its own best-effort hold: it covers its appliance's jumps until it runs dry, and refills when that appliance switches off. When several appliances want battery power at once, the available power is divided between them. The authors argue that "customized user control is key" — so set your own budget.
Your privacy budget Simulated day
What the meter reveals with this budget
Each group's big jumps over the simulated day: ■ hidden ■ delayed (shown later) ■ seen. Tap a row to inspect its events.
The algorithm must act before the next reading arrives, so it predicts it. The paper models a power trace as a Markov chain : the readings are grouped into a few typical power levels (states), and the chance of moving from one state to another is counted from past data. The forecast for the next reading is simply the most likely next state. On the measured whole-home data, this one-step forecast tracked the real trace closely (the paper's Fig. 3).
4 · Measuring privacy
Three detectives, three kinds of evidence
How do you score whether a jump is hidden? The paper compares the real jumps (dp) with the jumps the meter shows (dp′) using three yardsticks. Each attacks the problem from a different angle — and they can disagree.
Choose one, then try the sliders.
Variational distance ½ Σ |P₁ − P₂| between how often each jump size occurs. 0 = identical.
Cluster similarity: re-learns jump-size classes for each trace and counts real jumps whose class has gone missing.
Regression: fits a line between real and metered jumps, moment by moment. Protection = 1 − R².
Measured The selective algorithm did exactly this to the washer (meter 7). Cluster similarity rated only 33.33% (B1) and 14.29% (B2) of its jumps hidden, while regression rated protection at 94.49% and 99.99% : the battery mostly shifted the washer's jumps in time rather than erasing them. As the authors put it, several metrics "allow us to 'attack' the algorithm in manifold ways": the first two judge which events happen, regression also judges when.
The sorting detective is the hardest to fool, in the measured data too: cluster similarity gave the lowest scores in the paper's results (only 1.36% hidden for the whole home under the selective algorithm with B1), while regression often exceeded 90%. On this page the sorting detective compares how many jumps fall in each size class regardless of timing, and the timing detective fits a straight line between real and metered jumps — simplified versions of the paper's methods.
5 · The price of privacy
Hiding everything wears the battery out
Every charge and discharge costs something: energy losses and battery wear. The authors compared the two algorithms on three measures — how much energy flows out of the battery per day ( battery energy throughput ), how much the cycling shortens battery life ( cycle-life efficiency , where 1 is best), and how peaky the metered load is for the grid ( peak-to-average power ratio , where lower means smoother demand).
Battery charge through the day Simulated
Measured over 13 days Table V
6 · The evidence
What 13 days of real data showed
Where the energy went Table I
Average daily consumption: 2,930.2 Wh. Hatched: fridge and others, which the selective algorithm leaves unprotected (59.92%).
Privacy scores per meter Table IV
The headline results. For the home as a whole, the hide-everything algorithm disguised more (variational distance 0.38 versus 0.11 for selective), because the selective algorithm leaves the fridge and background loads alone by design. But the selective algorithm still protected individual appliances well — small quotas were enough because each appliance's events are sparse, so its share has time to refill — and it did so with about a third of the battery energy throughput and gentler charge cycles. The prediction step worked too: consumption events "can be predicted quite well".
The battery behaves like a buffer in a data network: over the 13 days the energy drawn from the grid hardly changed (38,092 Wh with no battery; between 37,736 and 38,059 Wh with one). The battery rearranges when energy arrives, not how much is used.
7 · Back to the kettle
A home that decides what its meter says
The authors conclude that home energy management can reduce how much sensitive usage information leaks, and that a battery is used more effectively when protection is selective. In their words, "the user may wish to hide the use of kettles, but not the use of a washing machine".
They also sketch what could come next: smarter privacy algorithms that emulate appliances such as a TV or lights, to mask unusually long periods when nobody is home; studies of how many homes doing this at once would affect the grid; and balancing privacy against energy use and cost. And they are clear that ElecPrivacy is not a substitute for rules on how meter data is collected and used — the two should work together.
Source and notes
G. Kalogridis, R. Cepeda, S. Z. Denic, T. Lewis and C. Efthymiou, “ElecPrivacy: Evaluating the Privacy Protection of Electricity Management Algorithms,” IEEE Transactions on Smart Grid, vol. 2, no. 4, pp. 750–758, December 2011. Toshiba Research Europe, Telecommunications Research Laboratory, Bristol, UK. doi:10.1109/TSG.2011.2160975
All visualisations on this page are original and drawn in code; no figures from the paper are reproduced. Scenes marked Simulated use a synthetic day whose appliance groups follow the paper's Table I; power shapes, wattages and battery behaviour are simplified re-implementations of the algorithms the paper describes, with next-step demand treated as known. Values marked Measured or "from the paper" are quoted from its Tables I, II, IV and V and text.
Topics in this iPaper
Engineering
Smart metering privacy, energy management, rechargeable batteries, power routing.
Abstract and Index Terms, page 750
Computer science
we propose one well known information theoretic quantity which can be used to compare two sources of information, the variational distance (VD)
Section IV-A Variational Distance, page 754
smart meters
The data collected by a home smart meter
Abstract, page 750
data privacy
In this paper, we study how home energy resources can be used to protect the privacy of the collected data.
Abstract, page 750
energy storage
rechargeable battery resources can help protect privacy of particular appliances
Section V Conclusion, page 757
markov chains
We shall demonstrate how a simple Markov chain model can be applied in representing a PLS.
Section III-B Modeling Power Load Signature by Markov Chains, page 753
cluster analysis
Cluster techniques take a set of data with a distance metric
Section IV-B Cluster Similarity, page 754