path: root/paper/safety-reset-paper.tex

\documentclass[letterpaper,twocolumn,10pt]{article}
\usepackage{usenix}

\usepackage[T1]{fontenc}
\usepackage[
    backend=biber,
    style=numeric,
    natbib=true,
    url=false, 
    doi=true,
    eprint=false
    ]{biblatex}
\addbibresource{safety-reset.bib}
\usepackage{amssymb,amsmath}
\usepackage{eurosym}
\usepackage{wasysym}

\usepackage[binary-units]{siunitx}
\DeclareSIUnit{\baud}{Bd}
\DeclareSIUnit{\year}{a}
\usepackage{commath}
\usepackage{graphicx,color}
\usepackage{subcaption}
\usepackage{array}
\usepackage{hyperref}
\usepackage{enumitem}

\renewcommand{\floatpagefraction}{.8}
\newcommand{\degree}{\ensuremath{^\circ}}
\newcolumntype{P}[1]{>{\centering\arraybackslash}p{#1}}
\newcommand{\partnum}[1]{\texttt{#1}}

\begin{document}

% https://eepublicdownloads.entsoe.eu/clean-documents/pre2015/publications/entsoe/Operation_Handbook/Policy_1_Appendix%20_final.pdf

\date{}
\title{Ripples in the Pond: Transmitting Information through Grid Frequency Modulation}
\author{Jan Sebastian Götte \and Liran Katzir \and Björn Scheuermann}
%\institute{TU Darmstadt\\ Communication Networks Lab\\ \email{safetyreset@jaseg.de}
%\and Tel Aviv University\\ Faculty of Engineering\\ \email{lirankat@tau.ac.il}
%\and TU Darmstadt\\ Communication Networks Lab\\ \email{scheuermann@informatik.hu-berlin.de}}
\maketitle
%\keywords{Security, privacy and resilience in critical infrastructures \and Security and privacy in ``internet of
%things'' \and Cyber-physical systems \and Hardware security \and Network Security \and Energy systems \and Signal theory}

\begin{abstract}
    Previous work has explored the scenario of an attacker compromising a large number of Smart Meters that are equipped
    with remote disconnect switches, and using these remote-controllable switches to cause a large-scale outage.
    Previous work focuses on attack prevention. In this paper, we will instead look at recovery after a successful
    attack. To transmission system operators (TSOs), the major challenge after such a Smart Meter-triggered outage is
    that the attacker will likely persist through the outage, and compromised Smart Meters will resume malicious
    activity after their power is restored. In the event of such an attack, TSOs would need a way to remotely put these
    compromised devices into a \emph{safe} mode of operation. 

    Given that public telecommunications networks including the internet, cellular networks, and LoRa base stations may
    also be disrupted during a large-scale blackout, the challenging aspect of this remote \emph{Safety Reset} is the
    communication channel between TSO and the smart meter. For this purpose, in this paper we propose a simple yet
    effective communication channel based on modulating grid frequency by modulating the power of a connected load or
    generator. Our proposed communciation channel (1) requires minimal infrastructure, (2) has a reach spanning the
    entire power grid and (3) is fully independent of other telecommunication networks and functions even under severe
    disruption of the grid.
\end{abstract}

\section{Introduction}

With the rollout of the smart grid, the IT security of electrical infrastructure has attracted increased attention in
the last years. Smart Grid security has two major components: The security of central SCADA systems, and the security
of equipment at the consumer premises such as smart meters and IoT devices. While there is previous work on both sides,
their interactions have not yet received much attention.

In this paper, we consider the previously proposed scenario where a large number of compromised consumer devices is used
alone or in conjunction with an attack on the grid's central SCADA systems to destabilize the grid by rapidly modulating
the total connected load. Previous work considered compromised smart meters with integrated remote disconnect switches
as likely candidates for such an attack, but the same attack can also be performed using compromised IoT devices.  Such
attacks are hard to mitigate, and existing literature focuses on hardening device firmware to prevent compromise.
Despite the infeasibility of perfect firmware security, there is little research on \emph{post-compromise} mitigation
approaches. A core issue with post-attack mitigation is that the devices normal network connection may not work due to
the attack and as such an out-of-band communication channel is necessary.

We propose a \emph{safety reset} controller that is controlled through a novel, resilient, grid-wide powerline
communication technique. Our safety reset controller can be fitted into any Smart Meter or IoT device. Its purpose is to
await an out-of-band command to put the device into a safe state (e.g. \emph{relay on} or \emph{light on}) that
interrupts attacker control over the device. The safety reset controller is separated from the system's main application
controller and does not have any conventional network connections to reduce attack surface and cost.

We propose a resilient grid-wide broadcast channel based on modulating grid frequency. This channel can be operated by
transmission system operators (TSOs) even during black-start recovery procedures and in this situation bridges the gap
between the TSO's private network and the consumer devices. To demonstrate our proposed channel, we have implemented a
system that transmits error-corrected and cryptographically secured commands.

Our approach differs from traditional Powerline Communication (PLC) systems in that it reaches every device within one
synchronous area as the signal is embedded into the fundamental grid frequency. Traditional PLC uses a superimposed
voltage, which is quickly attenuated across long distances.

\begin{figure}
    \centering
    \includegraphics[width=0.4\textwidth]{flowchart}
    \caption{Structural overview of our concept. 1 - Government authority or utility operations center. 2 - Emergency
    radio link. 3 - Aluminium smelter. 4 - Electrical grid. 5 - Target smart meter.}
    \label{fig_intro_flowchart}
\end{figure}

Figure~\ref{fig_intro_flowchart} shows an overview of our concept.  Two scenarios for its application are before or
during a cyberattack, to stop an attack on the electrical grid in its tracks, and after an attack while power is being
restored to prevent a repeated attack. In both scenarios, our concept is fully independent of all public communication
networks (such as the Internet or mobile networks) as well as broadcast systems (such as cable television or terrestrial
broadcast radio). A grid frequency-based system can function as long as power is still available, or as soon as power is
restored after the attack. One powerful function this allows is ``flushing out`` an attacker from compromised smart
meters after an attack, before restoring smart meter internet connectivity.

Using simulations we have determined that control of a $\SI{25}{\mega\watt}$ load such as a large aluminium smelter,
load bank or photovoltaic farm would allow for the transmission of a crytographically secured \emph{reset} signal within
$15$ minutes. We have designed and constructed a proof-of-concept prototype receiver that demonstrates the feasibility
of decoding such signals on a resource-constrained microcontroller.

\subsection{Motivation}

Consumer devices are increasingly becoming \emph{smart}. Large numbers of IoT devices are connected through the public
internet, and in several countries internet-connected Smart Meters can disconnect entire households from the grid in
case of unpaid bills. The increasing proliferation of smart devices on the consumer side presents an opportunity to grid
operators, who rely on forecasts for the cost-optimized control of generation and power flow. The core of the
\emph{Smart Grid} vision is that utilities can now gather detailed data for more accurate consumption forecasts, and in
some cases can even adjust parameters of large devices like water heaters to smooth out load spikes.

However, this increased degree of visibility and control comes with an increased IT security risk. In this paper we
focus on scenarios where an attacker compromises a large number of grid-connected remote-controllable devices. This may
be simple smart home devices such as IoT light bulbs, but it may also include Smart Meters that are outfitted with a
remote disconnect switch as is common in some countries. By rapidly switching large numbers of such devices in a
coordinated manner, the attacker has the opportunity to de-stabilize the electrical grid. % FIXME citation

Previous work on IoT and Smart Grid security has focused on the prevention of attacks though firmware security measures.
While research on prevention is undoubtably important, we estimate that its practical impact will be limited by the vast
diversity of implementations found in the field combined with the slow update cycles inherent to non-functional firmware
enhancements for consumer devices. We predict that it would be a Sisyphean task to secure sufficiently many devices
to deny an attacker the critical mass needed to cause trouble. For this reason, in this paper we focus on recovery after
an attack.

\subsection{Black-start recovery}

The recovery from a large-scale power outage is a complex operational challenge. Large outages are caused by cascading
failures. Since all consumers and producers that are connected to the electrical grid are physically coupled through the
electromotive force, a fault in one part of the grid affects all devices connected across the grid. To function, the
grid relies on a delicate balance between electricity generation, transmission and consumption.  When this balance is
disturbed, cascading failures can occur. A transmission line shutting off can lead other, nearby lines to overload and
shut off. Due to the electromechanical coupling of all machines connected to the grid, a generator or consumer suddenly
shutting off causes a transient in the grid's frequency. If the frequency goes too far out of bounds, protection devices
take power plants and large industrial loads offline.

The recovery from a large-scale outage requires the grid's operators to bring generators and loads back online one by
one while continuously maintaining balance between generation and consumption to avoid their protection devices shutting
them down again. To coordinate this process, transmission system operators cannot rely on the public internet or
cellular networks, as they may not work during a large-scale power outage. Instead, they maintain private communication
infrastructure using dedicated lines rented from telecommunciations providers, fibers run along transmission lines, and
dedicated radio links.

To start from a complete outage, first a number of \emph{black start}-capable power stations that can start by
themselves without any external power are brought online. With their help, other power stations and consumers are
gradually brought online until a part of the grid has been restored to nominal operation. This process can be performed
simultaneously in different parts of the grid. After these \emph{islands} have been restored, they can then be joined to
restore the grid to its normal state.

\subsection{Contents}

Starting from a high level architecture, we have carried out simulations of our concept's performance under real-world
conditions. Based on these simulations we implemented an end-to-end prototype of our proposed safety reset controller as
part of a realistic smart meter demonstrator. Finally, we experimentally validated our results and we will conclude with
an outline of further steps towards a practical implementation.

This work contains the following contributions:
\begin{enumerate}[topsep=4pt]
    \item We introduce Grid Frequency Modulation (GFM) as a communication primitive. % FIXME done before in that one paper
    \item We elaborate the fundamental physics underlying GFM and theorize on the constrains of a practical
        implementation.
    \item We design a communication system based on GFM.
    \item We carry out extensive simulations of our systems to determine its performance characteristics.
\end{enumerate}

\subsection{Notation}

To a computer scientist there is one confusing aspect to the theory of grid frequency modulation. GFM can be seen as a
frequency modulation (FM) with a baseband signal in the band below approximately $f_m = \SI{5}{\hertz}$ that is
modulated on top of a carrier signal at $f_c = \SI{50}{\hertz}$ in case of the European electrical grid. The frequency
deviation $f_\Delta$ that the modulated carrier deviates from its nominal value of $f_m$ is very small at only a few
milli-Hertz.

When grid frequency is measured by first digitizing the mains voltage waveform, then de-modulating digitally, the FM's
SNR is very high and is dominated by the ADC's quantization noise and nearby mains voltage noise sources such as
resistive droop due to large inrush current of nearby machines.

Note that both the carrier signal at $f_c$ and the modulation signal at $f_m$ both have unit Hertz. To disambiguate
them, in this paper we will use \textbf{bold} letters to refer to the carrier waveform $\mathbf{U}$ or frequency
$\mathbf{f_c}$ as well as its deviation $\mathbf{f_\Delta}$, and we will use normal weight for the actual modulation
signal and its properties such as $f_m$.

\section{Related work} 
\label{sec_related_work}

Previous work has analyzed Smart Grid security from numerous angles and made several suggestions towards its
improvement.  Apart from the critical location that Smart Grid devices occupy, they are computer systems like many
others.  Thus, for IT security purposes the Smart Grid is simply an aggregation of embedded control and measurement
devices that are part of a large control system. These devices share the same security concerns that apply to embedded
systems in general. 

\subsection{Smart Meter Security}

Where programmers have been struggling for decades now with issues such as input validation~\cite{leveson01}, the same
potential issue raises security concerns in smart grid scenarios as well~\cite{mo01, lee01}.  Only, in smart grid we
have two complicating factors present: many components are embedded systems, and as such inherently hard to update.
Also, the smart grid and its control algorithms act as a large (partially) distributed system making problems such as
input validation or authentication harder~\cite{blaze01} and adding a host of distributed systems problems on
top~\cite{lamport01}.

Given that the electrical grid is essential infrastructure, these issues are significant. Attacks on the electrical grid
may have grave consequences~\cite{anderson01,lee01} while the long replacement cycles of various components make the
system slow to adapt. Thus, components for the smart grid need to be built to a higher standard of security than e.g.\
IoT devices to live up to well-funded attackers decades down the road.  Another implication of their long service life
is that their agility w.r.t.\ post-hoc mitigations through firmware updates is limited.

%Another fundamental challenge in smart grid implementations is the central role of smart electricity meters in the
%smart grid ecosystem. Smart meters are used both for highly-granular load measurement and in some countries also for
%load switching~\cite{zheng01}.
Smart electricity meters are effectively consumer devices built down to a certain price point. The small market served
by a single smart meter implementation limits how much effort a vendor can spend on firmware security. Landis+Gyr, a
large manufacturer that makes most of its revenue from utility meters state in their 2019 annual report that they
invested \SI{36}{\percent} of their total R\&D budget on embedded software while spending only \SI{24}{\percent} on
hardware R\&D~\cite{landisgyr01,landisgyr02}, indicating significant tension between firmware security and the vendor's
bottom line. 

% FIXME more sources!

\subsection{The state of the art in embedded security}

Embedded software security generally is much harder than security of higher-level systems. The primary two factors
affecting this are that on one hand, embedded devices usually run highly customized firmware that (often by necessity)
is rarely updated. On the other hand, embedded devices often lack advanced security mechanisms such as memory management
units that are found in most higher-power devices. Even well-funded companies continue to have trouble securing their
embedded systems. A spectacular example of this difficulty is the 2019 flaw in Apple's iPhone SoC first-stage ROM
bootloader that allows for the full compromise of any iPhone before the iPhone X given physical access to the
device~\cite{heise01}. iPhone 8, one of the affected models, was still being manufactured and sold by Apple until April
2020.  In another instance in 2016, researchers found multiple flaws in Samsung's implementation of ARM TrustZone
``secure world'' firmware that Samsung used for their own mobile phone SoCs.  The flaws they found were both severe
architectural flaws such as secret user input being passed through untrusted userspace processes without any protection
as well as shocking cryptographic flaws such as
CVE-2016-1919\footnote{\url{http://cve.circl.lu/cve/CVE-2016-1919}}~\cite{kanonov01}.  And Samsung is not the only large
multinational corporation having trouble securing their secure firmware implementation. In 2014 researchers found an
embarrassing integer overflow flaw in the low-level code handling untrusted input in Qualcomm's QSEE
firmware~\cite{rosenberg01}. For an overview of ARM TrustZone including a survey of academic work and past security
vulnerabilities of TrustZone-based firmware see~\cite{pinto01}.

If even companies with R\&D budgets that rival some countries' national budgets at mass-market consumer devices
have trouble securing their mass market secure embedded software stacks, what is a much smaller smart meter manufacturer
to do?  Especially if national standards mandate complex protocols such as TLS that are tricky to implement
correctly~\cite{georgiev01}, this manufacturer will be short on options to secure their product.

\subsection{Attack surface in the smart grid}

From the incidents we outlined in the previous paragraphs we conclude that in smart metering technology, market
incentives do not currently provide the conditions for a level of device security that will reliably last the coming
decades. Considering this tension, in this paragraph we examine the cyberphysical risks that arise from attacks on the
smart grid in the first place. These risks arise at three different infrastructure levels.

The first level is that of attacks on centralized control systems. This type of attack is often cited in popular
discourse and to our knowledge is the only type of attack against an electric grid that has ever been carried out in
practice at scale~\cite{lee01}. Despite their severity, these attacks do not pose a strictly \emph{scientific} challenge
since they are generic to any industrial control system. Their causes and countermeasures are generally well-understood
and the hardest challenge in their prevention is likely to be budgetary constraints.

Beyond the centralized control systems, the next target for an attacker may be the communication links between those
control systems and other smart grid components. While in some countries such as Italy special-purpose systems such as
PLC are common~\cite{ec03}, overall, IP-based technologies have proliferated according to the larger trend towards
IP-based communications.  This proliferation of IP links brings along the possibility for the application of generic
network security measures from the IP world to the smart grid domain. In this way, a standardized, IP-based protocol
stack unlocks decades of network security improvements at little cost.

Beyond these layers towards the core of the smart grid's control infrastructure, an attacker might also corrupt the
network from the edges and target the endpoint devices itself. The large scale deployment of networked smart meters
creates an environment that is favorable to such attacks.
% FIXME cite RECESSIM landis+gyr protocol hacking wiki/youtube

\subsection{Cyberphysical threats in the smart grid}

Assuming that an attacker has compromised devices on any of these levels of smart grid infrastructure, what could they
do with their newly gained power? The obvious action would be to switch off everything. Of all scenarios,
this is both the most likely in practice---it is exactly what happened in the Russian cyberattacks on the Ukranian
grid~\cite{lee01}---but it is also the easiest to mitigate since the vulnerable components are few and centralized.
Mitigations include the installation of fail safes as well as a defense in depth approach to hardening the grid's
cyber infrastructure.

Another possible action for an attacker would be to forge energy measurements in an attempt to cause financial mayhem.
Both individual consumers as well as the utility could be targeted by such an attack. While such an attack might have
localized success, larger-scale discrepancies will likely quickly be caught by monitoring systems. For example, if a
large number of meters in an area systematically under- or over-reported their energy readings, meter readings across
the affected area would no longer add up with those of monitoring devices in other locations in the transmission and
distribution grid.

In some countries, smart meter functionality goes beyond mere monitoring devices and also includes remotely controlled
switches. There are two types of these switches: Switches to support \emph{Demand-Side Management} (DMS) and cut
off-switches that are used to punish defaulting customers. Demand Side Management is when a grid operator can remotely
control the timing of large, non-time-critical loads on the customer's premises~\cite{dzung01}. A typical example of this
is a customer using an electric water heater: The heater is outfitted with a large hot water storage tank and is
connected hooked up to the utility's DSM system. The customer does not care when exactly their water is heated as long
as there is enough of it, and the utility offers them cheaper rates for the electricity used for heating in exchange for
control over its precise timing.  The utility uses this control to even out peaks in the consumption/production
imbalance, remotely enabling DSM systems during off-peak times and disabling them during peak hours.  In contrast to
DSM, cut-off switches are switches placed in between the grid and the entire customer's household such that the utility
can disconnect non-paying customers without incurring the expense of sending a technician to the customer's premises.
Unlike DSM systems, cut-off switches are not opt-in~\cite{anderson01,temple01}.  An attack that uses cut-off switches
would obviously immediately cause severe mayhem. Attacks on DSM may have more limited immediate impact as affected
consumers may not notice an interruption for several hours.

Instead of switching off loads outright, an attack employing DSM switches (and potentially also cut-off switches) could
choose to target the grid's stability.  By synchronizing many compromised smart meters to switch on and off a large
load capacity, an attacker might cause the entire electrical grid to oscillate~\cite{kosut01,wu01,kim01}. As a large
system of coupled mechanical systems, the electrical grid exhibits a complex frequency-domain behavior.  Resonance
effects, colloquially called ``modes'', are well-studied in power system
engineering~\cite{rogers01,grebe01,entsoe01,crastan03}. As they can cause issues even under normal operating conditions,
a large effort is invested in dampening these resonances. Howewer, fully eliminating them under changing load conditions
may not be achievable.

\subsection{Communication Channels on the Grid}

A core part of intervening with any such cyberattack is the ability to communicate remediary actions to the devices
under attack.  There is a number of well-established technologies for communication on or along power lines. We can
distinguish three basic system categories: systems using separate wires (such as DSL over landline telephone wiring),
wireless radio systems (such as LTE) and \emph{Power Line Communication} (PLC) systems that reuse the existing mains
wiring and superimpose data transmissions onto the 50 Hz mains sine~\cite{gungor01,kabalci01}.

During a large-scale cyberattack, availability of internet and cellular connectivity cannot be relied upon. An attacker
may already have disabled such systems in a separate attack, or they may go down along with parts of the electrical
grid. Traditional powerline communication systems or an utitly's proprietary wireless systems would work, but at a range
of no more than several tens of kilometers reaching all meters in a country would require a large upfront infrastructure
investment.

\section{Grid Frequency as a Communication Channel}

We propose to approach the problem of broadcasting an emergency signal to all smart meters within a synchronous area by
using grid frequency as a communication channel.  Despite the awesome complexity of large power grids, the physics
underlying their response to changes in load and generation is surprisingly simple. Individual machines (loads and
generators) can be approximated by a small number of differential equations and the entire grid can be modelled by
aggregating these approximations into a large system of nonlinear differential equations. As a consequence, small signal
changes in generation/consumption power balance cause an approximately proportional change in
frequency~\cite{kundur01,crastan03,entsoe02,entsoe04}.  This \emph{Power Frequency Charactersistic} is about
\SI{25}{\giga\watt\per\hertz} for the continental European synchronous area according to European electricity grid
authority ENTSO-E.

If we modulate the power consumption of a large load such as a multi-megawatt aluminium smelter, this modulation will
result in a small change in frequency according to this characteristic. As long as we stay within the operational limits
set by ENTSO-E~\cite{entsoe02,entsoe03}, this change will not degrade the operation of other parts of the grid. The
advantages of grid frequency modulation are the fact that a single transmitter can cover an entire synchronous area as
well as low receiver hardware complexity.

To the best of the authors' knowledge, grid frequency modulation has only ever been proposed as a communication channel
at very small scales in microgrids before~\cite{urtasun01} and has not yet been considered for large-scale application.

Compared to traditional channels such as DSL, LTE or LoraWAN, grid frequency as a communication channel has a large
resiliency advantage: If there is power, a grid frequency modulation system is operational. Both DSL and LTE systems not
only require power but also require large amounts of centralized infrastructure to operate. Mesh networks such as
LoraWAN can cover short distances up to $\SI{20}{\kilo\meter}$ without requiring infrastructure to be available, but for
longer distances LoraWAN relies on the public internet for its network backbone. Additionally, systems such as DSL, LTE
and LoraWAN are built around a point-to-point communication model and usually do not support a generic broadcast
primitive. During times when a large number of devices must be reached simultaneously this can lead to congestion of
local cellular towers or gateways.
Therefore, during an ongoing cyberattack, grid frequency is promising as a communication channel as only a single
transmitter facility must be operational for it to function, and this single transmitter can reach all connected devices
simultaneously. After a power outage, it can function as soon as electrical power is restored, even while the public
internet and mobile networks are still offline and it is unaffected by cyberattacks that target telecommunication
networks.

\subsection{Characterizing Grid Frequency}
\label{grid-freq-characterization}

In utility SCADA systems, Phasor Measurement Units (PMUs, also called \emph{synchrophasors}) are used to precisely
measure grid frequency among other parameters.  This task is much more complicated in practice than it might appear at
first glance since a PMU has to make extremely precise measurements, track fast changes in frequency and handle even
distorted input signals.  Detail on the inner workings of commercial phasor measurement units is scarce but there is a
large amount of academic research on sophisticated phasor measurement
algorithms~\cite{narduzzi01,derviskadic01,belega01}.

Since we do not need reference standard-grade accuracy for our application we chose to start with a very basic algorithm
based on short-time fourier transform (STFT). Our system uses the universal frequency estimation approach of
experimental physicists Gasior and Gonzalez at CERN~\cite{gasior01}.  The Gasior and Gonzalez algorithm~\cite{gasior01}
passes the windowed input signal through a DFT, then interpolates the signal's fundamental frequency by fitting a
wavelet such as a Gaussian to the largest peak in the DFT results. The bias parameter of this curve fit is an accurate
estimation of the signal's fundamental frequency. This algorithm is similar to the simpler interpolated DFT algorithm
used as a reference in much of the phasor measurement literature~\cite{borkowski01}.

To collect ground truth measurements for our analysis of grid frequency as a communication channel, we developed a device
to safely record real mains voltage waveforms.  Our system consists of an \texttt{STM32F030F4P6} ARM Cortex M0
microcontroller that records mains voltage using its internal 12-bit ADC and transmits measured values through a
galvanically isolated USB/serial bridge to a host computer. We derive our system's sampling clock from a crystal oven to
avoid frequency measurement noise due to thermal drift of a regular crystal: \SI{1}{ppm} of crystal drift would cause a
grid frequency error of $\SI{50}{\micro\hertz}$. We validated the performance of our crystal oven solution by
benchmarking it against a GPS 1pps reference.

% FIXME measurement results, spectra

\section{Grid Frequency Modulation}

Given the grid characteristics we measured using our custom waveform recorder and a model of our transmitter, we can
derive parameters for the modulation of our broadcast system.  In its most basic form a transmitter for grid frequency
modulation would be a very large controllable load connected to the power grid at a suitable vantage point. A spool of
wire submerged in a body of cooling liquid such as a small lake along with a thyristor rectifier bank would likely
suffice to perform this function during occasional cybersecurity incidents.  We can however decrease hardware and
maintenance investment even compared to this rather uncultivated solution by repurposing large industrial loads
as transmitters.  Going through a list of energy-intensive industries in Europe~\cite{ec01}, we found that an aluminium
smelter would be a good candidate.  In aluminium smelting, aluminium is electrolytically extracted from alumina solution.
High-voltage mains power is transformed, rectified and fed into about 100 series-connected electrolytic cells forming a
\emph{potline}. Inside these pots alumina is dissolved in molten cryolite electrolyte at about \SI{1000}{\degreeCelsius}
and electrolysis is performed using a current of tens or hundreds of Kiloampère. The resulting pure aluminium settles at
the bottom of the cell and is tapped off for further processing.

Aluminium smelters are operated around the clock, and due to the high financial stakes their behavior under power
outages has been carefully characterized by the industry. Power outages of tens of minutes up to two hours reportedly do
not cause problems in aluminium potlines~\cite{eisma01,oye01}. Recently, even techniques for intentional power modulation
without affecting cell lifetime or product quality have been developed to take advantage of variable energy
prices.~\cite{duessel01,eisma01,depree01}.  An aluminium plant's power supply is controlled to constantly keep all
smelter cells under optimal operating conditions. Modern power supply systems employ large banks of diodes or SCRs to
rectify low-voltage AC to DC to be fed into the potline~\cite{ayoub01}. Potline voltage is controlled through a
combination of a tap changer and a transductor.  Individual cell voltages are controlled by changing the physical
distance between anode and cathode distance.  In this setup, power can be modulated fully electronically. Since this
system does not have any mechanical inertia, high modulation rates can reasonably be achieved.

In~\cite{depree01}, the authors describe a setup where a large Aluminium smelter in continental Europe is used as
primary control reserve for frequency \emph{regulation}. In this setup, a rise time of $\SI{15}{\second}$ was achieved
to meet the $\SI{30}{\second}$ requirement posed by local standards for primary control. In their conclusion, the
authors note that for their system, an energy storage capacity of $\SI{7.7}{\giga\watt\hour}$ is possible if all plants
of a single operator are used. Given the maximum modulation depth of $\SI{100}{\percent}$ for up to one hour that is
mentioned by the authors, this results in an effective modulation power of $\SI{7.7}{\giga\watt}$. Over a longer
timespan of $\SI{48}{\hour}$, they have demonstrated a $\SI{33}{\percent}$ modulation depth which would correspond to
a modulation power of $\SI{2.5}{\giga\watt}$.

From this brief literature review, we conclude that a modulation of part of an aluminium smelter's power consumption
most likely is possible at no significant production impact and low infrastructure cost (such as for shell heat
exchangers as used in~\cite{depree01}). Aluminium smelters are connected to the grid in a way that they do not pose a
danger to other nearby consumers when they turn off or on parts of the plant, as this is commonplace during routine
maintenance activities. They are very large consumers of electrical power, but they are still small when seen in
relation to the entire grid.

\subsection{Parametrizing Modulation for GFM}

Modulating $\SI{25}{\mega\watt}$ of smelter power would yield a frequency shift of $\SI{1}{\milli\hertz}$. At an RMS
frequency noise of around $\SI{10}{\milli\hertz}$ in the band around $\SI{1}{\hertz}$, this results in challenging SNR.
% FIXME properly calculate frequency noise density, SNR
Under such conditions, the obvious choice for modulation are spread-spectrum techniques. Thus, we approached the setting
using Direct Sequence Spread Spectrum for its simple implementation and good overall performance. DSSS chip timing
should be as fast as the transmitter's physics allow to exploit the low-noise region between
$\SI{0.2}{\hertz}$ to $\SI{2.0}{\hertz}$ in the frequency noise spectrum while avoiding any of the grid's oscillation modes. Going
past $\approx\SI{2}{\hertz}$ would put strain on the receiver's frequency measurement subsystem~\cite{belega01}. Using a
spread-spectrum technique allows us to reduce the effect of interference by spurious tones. In addition, spreading our
signal's energy over frequency also reduces the likelihood that we cause the grid to oscillate along any of its modes.

To test our proposed approach, we wrote a proof-of-concept modulator and demodulator in Python and tested this
proof-of-concept prototype with data captured from our grid frequency sensor. Our simulations covered a range of
parameters in modulation amplitude, DSSS sequence bit depth, chip duration and detection threshold.
Figure~\ref{fig_ser_nbits} shows symbol error rate (SER) as a function of modulation amplitude with Gold sequences of
several bit depths. As can be seen, realistic modulation amplitudes are in the range around $\SI{1}{\milli\hertz}$. In
the continental European synchronous area, this corresponds to a modulation power of approximately
$\SI{25}{\mega\watt}$.  Figure~\ref{fig_ser_thf} shows SER against detection threshold relative to background noise.
Figure~\ref{fig_ser_chip} shows SER against chip duration for a given fixed symbol length. As expected from looking at
our measured grid frequency noise spectrum, performance is best for short chip durations and worsens for longer chip
durations since shorter chip durations move our signals' bandwidth into the lower-noise region from $\SI{0.2}{\hertz}$
to $\SI{2}{\hertz}$.
%FIXME introduce term "chip" somewhere

\begin{figure}
    \centering
    \includegraphics[width=0.6\textwidth]{../notebooks/fig_out/dsss_gold_nbits_overview}
    \caption{Symbol Error Rate as a function of modulation amplitude for Gold sequences of several lengths.}
    \label{fig_ser_nbits}
\end{figure}

\begin{figure}
    \centering
    \hspace*{-1cm}\includegraphics[width=1.2\textwidth]{../notebooks/fig_out/dsss_thf_amplitude_5678}
    \caption{SER vs.\ Amplitude and detection threshold. Detection threshold is set as a factor of background noise
    level.}
    \label{fig_ser_thf}
\end{figure}

\begin{figure}
    \centering
    \hspace*{-1cm}\includegraphics[width=1.2\textwidth]{../notebooks/fig_out/chip_duration_sensitivity_6}
    \vspace*{-1cm}
    \caption{SER vs.\ DSSS chip duration.}
    \label{fig_ser_chip}
\end{figure}

\subsection{Parametrizing a proof-of-concept "Safety Reset" System Based on GFM}

Taking these modulation parameters as a starting point, we proceeded to create a proof-of-concept smart meter emergency
reset system. On top of the modulation described in the previous paragraphs we layered simple Reed-Solomon error
correction~\cite{mackay01} and some cryptography. The goal of our PoC cryptographic implementation was to allow the
sender of an emergency reset broadcast to authorize a reset command to all listening smart meters. An additional
constraint of our setting is that due to the extremely slow communication channel all messages should be kept as short
as possible. The solution we chose for our PoC is a simplistic hash chain using the approach from the Lamport and
Winternitz One-time Signature (OTS) schemes. Informally, the private key is a random bitstring. The public key is
generated by recursively applying a hash function to this key a number of times. Each smart meter reset command is then
authorized by disclosing subsequent elements of this series. Unwinding the hash chain from the public key at the end of
the chain towards the private key at its beginning, at each step a receiver can validate the current command by checking
that it corresponds to the previously unknown input of the current step of the hash chain. Replay attacks are prevented
by recording the most recent valid command. Keys revocation is supported by designating the last key in the chain as a
\emph{revocation key} upon whose reception the client devices advance their local hash ratchet without taking further
action.  This simple scheme does not afford much functionality but it results in very short messages and removes the
need for computationally expensive public key cryptography inside the smart meter.
% FIXME add more precise/formal description of crypto
% FIXME add description of targeting/scope function?
% FIXME somewhere above descirbe entire reset system architecture????!!!
% FIXME add description of disarm message (replay protection)

\subsection{Experimental results}

\begin{figure}
    \centering
    \includegraphics[width=0.6\textwidth]{prototype.jpg}
    \caption{The completed prototype setup. The board on the left is the safety reset microcontroller. It is connected
    to the smart meter in the middle through an adapter board. The top left contains a USB hub with debug interfaces to
    the reset microcontroller. The cables on the bottom left are the debug USB cable and the \SI{3.5}{\milli\meter}
    audio cable for the simulated mains voltage input.}
    \label{fig_proto_pic}
\end{figure}

For a realistic proof of concept, we decided to implement our signal processing chain from DSSS demodulator through
error correction up to our simple cryptography layer in microcontroller firmware and demonstrate this firmware on actual
smart meter hardware, shown in Figure~\ref{fig_proto_pic}. In our proof of concept a safety reset controller is
connected to the main application microcontroller of a smart meter. The reset controller is tasked with listening for
authenticated reset commands on the voltage waveform, and on reception of such a command resetting the smart meter
application controller by flashing a known-good firmware image to its memory.

The signal processing chain of our PoC is shown in Figure~\ref{fig_demo_sig_schema}. To interoperate with existing
implementations of SHA-512 and reed-solomon decoding, this implementation was written in the C programming language. To
demonstrate an application close to a field implementation, we chose an Easymeter \texttt{Q3DA1002} smart meter as our
reset target. This model is popular in the German market and readily available second-hand. The meter consists of three
isolated metering ASICs connected to a data logging and display PCB through infrared optical links. To demonstrate the
safety reset's firmware reset functionality, we connected our safety reset microcontroller to the Texas Instruments
\texttt{MSP430} microcontroller on the meter's display and data logging board through the JTAG debug interface that the
board's vendor had conveniently left accessible. We ported part of
\texttt{mspdebug}\footnote{\url{https://dlbeer.co.nz/mspdebug/}} to drive the meter microcontroller's JTAG interface and
wrote a piece of demonstrator code that overwrites the meter's firmware with one that displays an identifying string on
the meter's display after boot-up.

\begin{figure}
    \centering
    \includegraphics[width=\textwidth]{prototype_schema}
    \caption{The signal processing chain of our demonstrator.}
    \label{fig_demo_sig_schema}
\end{figure}

To measure grid frequency in our demonstrator, we ported the same code we used in
Section~\label{grid-freq-characterization} to our demonstrator, again using the voltage measured using the
microcontroller's internal ADC but using a regular crystal instead of a crystal oven for the microcontroller's system
clock.  Since we did not have an aluminium smelter ready, we decided to feed our proof-of-concept reset controller with
an emulated grid voltage sine wave from a computer's headphone jack. Where in a real application this microcontroller
would  take ADC readings of input mains voltage divided down by a long resistive divider chain, we instead feed the ADC
from a $\SI{3.5}{\milli\meter}$ audio input. For operational safety, we disconnected the meter microcontroller from its
grid-referenced capacitive dropper power supply and connected it to our reset controlller's debug USB power supply.

We performed several successful experiments using a signature truncated at 120 bit and a 5 bit DSSS sequence. Taking the
sign bit into account, the length of the encoded signature is 20 DSSS symbols. On top of this we used Reed-Solomon error
correction at a 2:1 ratio inflating total message length to 30 DSSS symbols. At the \SI{1}{\second} chip rate we used in
other simulations as well this equates to an overall transmission duration of approximately \SI{15}{\minute}. To give
the demodulator some time to settle and to produce more realistic conditions of signal reception we padded the modulated
signal unmodulated noise on both ends.

\section{Discussion}

For our proof of concept, before settling on the commercial smart meter we first tried to use an \texttt{EVM430-F6779}
smart meter evaluation kit made by Texas Instruments. This evaluation kit did not turn out well for two main reasons.
One, it shipped with half the case missing and no cover for the terminal blocks. Because of this some work was required
to get it electrically safe.  Even after mounting it in an electrically safe manner the safety reset controller
prototype would also have to be galvanically isolated to not pose an electrical safety risk since the main MCU is not
isolated from the grid and the JTAG port is also galvanically coupled. The second issue we ran into was that the
development board is based around a specific microcontroller from TI's \texttt{MSP430} series that is incompatible with
common JTAG programmers.

Our initial assumption that a development kit would be easier to program than a commercial meter did not prove to be
true. Contrary to our expectations the commercial meter had JTAG enabled allowing us to easily read out its stock
firmware without either reverse-engineering vendor firmware update files nor circumventing code protection measures.
The fact that its firmware was only available in its compiled binary form was not much of a hindrance as it proved not
to be too complex and all we wanted to know we found with just a few hours of digging in
Ghidra\footnote{\url{https://ghidra-sre.org/}}.

In the firmware development phase our approach of testing every module individually (e.g. DSSS demodulator, Reed-Solomon
decoder, grid frequency estimation) proved to be very useful. In particular debugging benefited greatly from being able
to run several thousand tests within seconds. In case of our DSSS demodulator, this modular testing and simulation
architecture allowed us to simulate thousands of runs of our implementation on test data and directly compare it to our
Jupyter/Python prototype. Since we spent more time polishing our embedded C implementation it turned out to perform
better than our Python prototype while still exhibiting the same fundamental response to changes to its parameters.

In accordance with our initial estimations we did not run into any code space nor computation bottlenecks for chosing
floating point emulation instead of porting over our algorithms to fixed point calculations. The extremely slow sampling
rate of our systems makes even heavyweight processing such as FFT or our brute force dynamic programming approach to
DSSS demodulation possible well within our performance constraints.

The safety reset controller does not require any peripherals except for an ADC. Thus we expect code size to be the main
factor affecting per-unit cost in an in-field deployment of our concept. At around \SI{64}{\kilo\byte}, our unoptimized
demonstrator firmware implementation is already on the lower end of the spectrum. Especially with some optimization we
expect safety reset controllers to be commercially viable given adequate political incentives.

\section{Conclusion}
\label{sec_conclusion} 

In this paper we have developed an end-to-end design of a reset system to restore smart meters to a safe operating state
during an ongoing large-scale cyberattack.  To allow our system to be triggered even in the middle of a cyberattack we
have developed a broadcast data transmission system based on intentional modulation of global grid frequency. We have
shown the viability of our end-to-end design through simulations. To put these simulations on a solid foundation we have
developed a grid frequency measurement methodology comprising of a custom-designed hardware device for electrically safe
data capture and a set of software tools to archive and process captured data. Our simulations show good behavior of our
broadcast communication system and give an indication that cooperating with a large consumer such as an aluminium smelter
would be a feasible way to set up a transmitter with low hardware overhead.  We have outlined a simple cryptographic
protocol ready for embedded implementation in resource-constrained systems that allows triggering a safety reset with a
response time of less than 30 minutes.  We have experimentally validated our system using simulated grid frequency data
in a demonstrator setup based on a commercial microcontroller as our safety reset controller and an off-the-shelf smart
meter. The next step in our evaluation will be to conduct an experimental evaluation of our modulation scheme in
collaboration with an utility and an operator of a multi-megawatt load.  Source code and electronics CAD designs are
available at the public repository listed at the end of this document.

\printbibliography[heading=bibintoc]

%%% FIXME remove appendix and work into text.

\center{
    \center{This is version \texttt{\input{version.tex}\unskip} of this paper, generated on \today. The git repository
    can be found at:}

    \center{\url{https://git.jaseg.de/safety-reset.git}}
}
\end{document}