\documentclass[11pt,a4paper,notitlepage,twoside]{report}
\usepackage[ngerman, english]{babel}
\usepackage[utf8]{inputenc}
\usepackage[a4paper, top=3cm, bottom=3.5cm, inner=3.5cm, outer=5cm, marginpar=3.8cm]{geometry}
\usepackage[T1]{fontenc}
\usepackage{amssymb}
\usepackage{amsmath}
\usepackage{listings}
\usepackage{eurosym}
\usepackage{wasysym}
\usepackage{extdash}
\usepackage{amsthm}
\usepackage{mwe}
\usepackage{tabularx}
\usepackage{multirow}
\usepackage{multicol}
\usepackage{tikz}
\usepackage{mathtools}
\usepackage{setspace}
\usepackage{titlesec}
\usepackage{fancybox}
\usepackage{fancyhdr}
\usepackage[binary-units,per-mode=fraction]{siunitx}
\usepackage[hidelinks]{hyperref}
\usepackage{commath}
\usepackage{graphicx,color}
\usepackage{ccicons}
\usepackage{subcaption}
\usepackage{float}
\usepackage{footmisc}
\usepackage{array}
\usepackage[underline=false]{pgf-umlsd}
\usetikzlibrary{calc}
\usepackage{epstopdf}
\usepackage{pdfpages}
\usepackage{etoolbox}
\usepackage{catchfile}
\usepackage{minitoc}
\usepackage{minted} % pygmentized source code
%\usepackage[pdftex]{graphicx,color}
%\usepackage{showframe} % Useful for page layout debugging

\DeclareSIUnit{\baud}{Bd}

\DeclarePairedDelimiter{\ceil}{\lceil}{\rceil}
\DeclarePairedDelimiter{\paren}{(}{)}

\usepackage[
    backend=biber,
    style=numeric,
    natbib=true,
    url=false, 
    doi=true,
    eprint=false,
    % Make the split online / other resource bibliographies behave
    defernumbers=true,
    ]{biblatex}
\addbibresource{../main.bib}
\DeclareSourcemap{
    \maps[datatype=bibtex]{
        \map{
            \step[fieldsource=doi,final]
            \step[fieldset=isbn,null]
            \step[fieldset=issn,null]
            \step[fieldset=url,null]
        }
        \map{
            \step[fieldsource=isbn,final]
            \step[fieldset=issn,null]
            \step[fieldset=url,null]
        }
    }
}

\renewcommand{\thesection}{\arabic{section}}
\renewcommand{\thesubsection}{\arabic{section}.\arabic{subsection}}
\renewcommand{\thesubsubsection}{\arabic{section}.\arabic{subsection}.\arabic{subsubsection}}

% Re-define heading formats to force single line spacing
\titleformat{\section}{\normalfont\large\bfseries\singlespacing}{\thesection}{1em}{}
\titleformat{\subsection}{\normalfont\large\bfseries\singlespacing}{\thesubsection}{1em}{}
\titleformat{\subsubsection}{\normalfont\large\bfseries\singlespacing}{\thesubsubsection}{1em}{}

\newcommand{\degree}{\ensuremath{^\circ}}
\newcolumntype{P}[1]{>{\centering\arraybackslash}p{#1}}
\definecolor{todoboxcolor}{RGB}{251 224 252}

\pagestyle{fancy}

\fancyhead[C]{}
\fancyhead[ER]{\footnotesize%
    \ifdefined\thesispreviewmode %
        (draft \texttt{\input{version.tex}\unskip}) %
    \fi %
\leftmark}
\fancyhead[OL]{\footnotesize\rightmark}
\fancyhead[EL,OR]{\thepage}

\fancyfoot[LCR]{}

\fancypagestyle{plain}{%
    \fancyhf{}%
    \renewcommand{\headrulewidth}{0pt}%
    \renewcommand{\footrulewidth}{0pt}%
}

\raggedbottom

\renewcommand{\chaptermark}[1]{\markboth{Chapter \thechapter: #1}{}}
\renewcommand{\sectionmark}[1]{\markright{\thesection\ #1}}
\addtolength{\headwidth}{\marginparsep}
\addtolength{\headwidth}{\marginparwidth}
\addtolength{\headwidth}{-1cm}

\newcommand{\todo}[1]{
    \ifdefined\thesispreviewmode
    \marginpar{
        \setlength{\fboxsep}{2mm}
        \shadowbox{
            \parbox{3cm}{
                \singlespacing
                \raggedright
                \textsf{
                    \small\textbf{To do}\\
                    \footnotesize#1
                }
            }
        }
    }
    \fi
}

\newcommand{\todoplaceholder}[1]{\textbf{TODO}\todo{#1}}

% https://tex.stackexchange.com/questions/30720/footnote-without-a-marker
\newcommand\blfootnote[1]{%
  \begingroup
  \renewcommand\thefootnote{}\footnote{#1}%
  \addtocounter{footnote}{-1}%
  \endgroup
}

\newcommand{\figurepath}{figures}
\graphicspath{{\figurepath}}
\newcommand{\figureattrib}[1]{%
    \input{\figurepath/#1.latex_meta} %
    \scriptsize
        \ifdefined\thesispreviewmode\resourcestate\ \resourcescale\\\fi%
        Resource: %
        \texttt{\resourcerepo/\resourcepath} %
        rev \texttt{\resourcerev} %
        (\underline{\href{\resourceurl}{link}})%
}

\newcommand{\draftgraphics}{\ifdefined\thesispreviewmode\textcolor{red}{\bfseries Not final graphics. }\fi}
\newcommand{\camerareadygraphics}{\ifdefined\thesispreviewmode Camera-ready graphics. \fi}
\newcommand{\scaledgraphics}[1]{\ifdefined\thesispreviewmode scaled-#1\else#1\fi}

\newcommand{\imgsource}[4]{\scriptsize%
    Image source: #1, #2 (\underline{\href{#4}{link}}). %
    Licensed #3.}

\hyphenation{a-me-na-ble}

\begin{document}
\dominitoc
\faketableofcontents

\chapter{Physical Security in Quantum Key Distribution}
\ifdefined\thesispreviewmode
{\Large \textbf{Draft build}, git revision \texttt{\input{version.tex}}}
\fi
\minitoc
\newpage
\setstretch{1.3}

\section{Cryptography in the Age of Quantum Computers}

For a decade or two now, Quantum Computing has been creating a buzz that nobody in Computer Science and adjacent fields
could evade. Originating in the 1980ies as a highly academic thought experiment applying ideas from Computer Science in
Quantum Physics, \todo{Add citation on QKD origins} its concepts have long found their way into popular science
articles. Quantum Computing encompasses a model of computation that is fundamentally different from the
\emph{classical}\footnote{ In Quantum Computing, the term \emph{classical} is used as the complement of \emph{quantum},
and refers to the digital computers we know and (sometimes) love. This terminology stems from the distinction between
classical and quantum physics.} digital circuits that underly all of modern computing. While at first this might seem
like a step backwards into the era of early 1900s analog computing,\todo{Add citation on early analog computing} the
capabilites of a future quantum computer promise to outpace those of any possible classical computer. Key to this
improved processing capability is a property called \emph{Quantum Parallelism}, referring to the fact that inside of a
quantum computer, a single \emph{quantum state} can simultaneously represent a multitude of states of a classical,
digital computer, encoded into a quantum \emph{superposition}. Furthermore, the quantum computer can operate on all
those states at once using a single \emph{quantum gate}.

The quantum gates of a quantum computer do not correspond directly to classical, digital logic. Applying Quantum
Parallelism to practical problems is more complicated than, simply translating a digital circuit that computes a
solution to a quantum circuit. Nevertheless, for certain problems \emph{quantum algorithms} have already been developed
that for large inputs promise to solve these problems much faster than any classical computer ever could. Two of these
algorithms, one by Shor and one by Grover \todo{Add citations on Shor's and Grover's algorithm} are what caused most of
the buzz around the field of quantum computing because they spell trouble for a large part of modern cryptography.
While neither is a threat under the current state of the art in quantum computing, assuming a sufficiently advanced
quantum computer both algorithms provide solutions to problems that are classically assumed to be \emph{hard} with
vastly improved asymptotical time complexity.

Besides the computational speed-up promised by Quantum Parallelism, there is one more interesting aspect of Quantum
Computing where it radically deviates from classical computing. The reason modern cryptography exists is that when we
transmit (or store!) classical information through some channel (or storage!) that we do not control, there is nothing
we can do to prevent an attacker from reading this information. Even with cryptography we cannot prevent this, but
cryptography gives us tools to very effectively make whatever information the attacker is able to read useless to them.

A basic principle of Quantum Physics is the \emph{No-Cloning Theorem}, which states that it is impossible to create an
identical, independent copy of an arbitrary, unknown quantum state. \todo{Add citation on No-Cloning Theorem}
An implication of this theorem is that when we encode classical information into quantum states in just the right way,
we can make it so that an attacker attempting to eavesdrop on our quantum information can only decode this information
by destroying the underlying quantum states it in the process, which can be detected statistically. This property can be
exploited to replace a number of classical asymmetric primitives in interactive settings, \todo{Add citation on
substitution, check if interactive only} the most popular application of which is replacing an asymmetric Diffie-Hellman
key exchange \todo{Add citation on DH-Kex} with a quantum process called Quantum Key Distribution (QKD) that yields much
of the same properties.

In the past decades, the field of cryptography has been fundamentally shaped by the development of Quantum Computing and
Quantum Key Distribution. However, the popular conception that all of today's cryptography will be broken and that we
have to start from scratch is not accurate. Quantum Computing poses an unique threat to modern cryptography, and Quantum
Key Distribution is a promising new tool, but the practical implications of both are much more subtle than how they are
often portrayed. In the remainder of this chapter, we will look into the practical implications of these quantum
technologies, and we will come to two major conclusions: First, that while the underlying cryptographic primitives will
change, apart from some engineering issues cryptography as a whole will remain largely the same. Second, that while
Quantum Key Distribution is hailed as a revolution for network security, its practical advantages will remain far short
of how it is usually conceptualized, and hardware security will assume a pivotal role in the practical security of
Quantum Key Distribution systems. The central role of hardware security in Quantum Key Distribution is a stark departure
from its relative irrelevance in today's applied cryptography.

Building on these conclusions, we will end this chapter with a study of a use case that illustrates a practical design
for a secure network employing Quantum Key Distribution. Relying on both established classical and quantum primitives
with known security properties we will elaborate how one can construct a large-scale network from those primitives
that uses IHSMs to provide practical security beyond the---surprisingly limited---extent of quantum security proofs.

\subsection{Computational Assumptions and Information\Hyphdash Theoretic Security}
\label{qc_comp_assum}

We have briefly mentioned that Quantum Computing promises to eventually provide a significant speed-up that can be
applied to solve many cryptographic problems fast enough for it to become a problem, but we have not elaborated on what
that means in practice. In this section, we will attempt convey a more concrete intuition of the magnitude of the threat
that both Shor's and Grover's algorithm and variants pose to modern cryptography.

\textcite{shorAlgorithmsQuantumComputation1994, shorPolynomialTimeAlgorithmsPrime1997} introduced several algorithms to
solve problems in polynomial time on a quantum computer that are still believed to be hard on classical computers today.
In the original conference paper and journal article, Shor introduces several algorithms based on a similar fundamental
approach. Depending on context, \emph{Shor's algorithm} usually refers to one of two of these algorithms that solve
integer factorization as used in RSA, and the discrete logarithm problem as used in the Diffie-Hellman key exchange,
respectively.

While Shor's algorithm attacks the foundations of most modern asymmetric cryptography, Grover's algorithm can be applied
to hash functionss and symmetric cryptography. Fundamentally, Grover's algorithm is a search algorithm that allows a
quantum computer to find one target entry out of an \emph{unstructured} list of $N$ source entries in
$\mathcal{O}\left(\sqrt{N}\right)$ time instead of the $\mathcal{O}\left(N\right)$ time that a classical computer would
require for an exhaustive search. Applied to cryptography, we model the key space of a symmetric cipher as the
unstructured list that is input to the algorithm, and set it to search for the key that results in the successful
decryption of a given ciphtertext.

An important nuance applying these algorithms to cryptography is that while both provide significant speed-ups over
classical computers, the speed-up of Shor's algorithm is exponential and effectively breaks most modern asymmetric
cryptography as it erases the asymmetric nature of the underlying mathematical problem's computational complexity. That
is, for an asymmetric cryptosystem susceptible to Shor's algorithm, there is no set of parameters that is large enough
to be safe.

In contrast to this, while Grover's algorithm radically speeds up the breaking of a symmetric cryptosystem, this
speed-up is only quadratic. In practice this means that it halves the security level \todo{definition, citation of
security level} of a given symmetric cipher. While this is bad news for applications that parameterize these symmetric
primitives to a security level at the lower end of what is considered secure today, the advantage provided by Grover's
algorithm can easily be compensated by doubling key size. Longer key sizes require more storage or bandwidth for the
additional bits and result in slightly slower operation of the cipher, but this additional cost is easily manageable
even without any improvement in today's hardware.

\textcite{impagliazzoPersonalViewAveragecase1995} provided a colloquial but useful analysis characterizing the
implications of which kinds of hard problems are solvable in practice, based on the observation that the fact that an
\emph{average} problem out of a class like $NP$ is solvable does not mean that most, or even many \emph{practical}
problems are solvable. \textcite{impagliazzoPersonalViewAveragecase1995} was published after Shor's algorithm was
discovered, and before Grover's algorithm was published. Impagliazzo foresaw that fast quantum algorithms could threaten
public key security, and their analysis remains relevant facing the outlook of quantum computing today.

Impagliazzo proposes a set of five scenarios that provide increasingly extensive computational hardness properies,
dubbed \emph{Algorithmica}, \emph{Heuristica}, \emph{Pessiland}, \emph{Minicrypt}, and \emph{Cryptomania}. In
Algorithmica, $P = NP$. In Heuristica, $P \ne NP$, but $NP$ problems are only intractable in the worst case, and
tractable on average. In Pessiland, problems exist that are hard on average, but there are no one-way functions and thus
there is no way to efficiently sample solved instances of hard problems.

The next scenario, Minicrypt is frequently cited in cryptographic works. In it, one-way functions exist, but there is no
public key cryptography. Minicrypt aligns well with a world in which fast quantum algorithms exist that solve the
computational problems underlying public key cryptosystems. Impagliazzo's last scenario is Cryptomania, which extends
Minicrypt with public key cryptography and aligns with the world view that is commonly assumed in cryptography today.

In Minicrypt, we assume that all computational problems that are amenable to public key cryptography fall. However, it is
not specified \emph{how} specifically this fall will happen---whether it will be classically, or by quantum
algorithms---leading to two sub-variants of the Minicrypt scenario. The pessimistic sub-variant is one where classical
algorithms solving all those problems are discovered. This scenario leads to identical conclusions to those Impagliazzo
drew. However, if we base our Minicrypt assumption instead on the availability of \emph{quantum } algorithms for these
problems, and thus on quantum computers being both powerful enough and generally available, we end up with an
interesting spin on the original Minicrypt scenario that recently has garnered some academic attention, receiving the
name Mini\textbf{Q}Crypt\cite{griloObliviousTransferMiniQCrypt2021, barootiPublicKeyEncryptionQuantum2023}. In
MiniQCrypt, on one hand, conventional public key cryptography is broken by quantum computers running Shor's algorithm,
but the key observation is that on the other hand, we can then use those quantum computers to do \emph{quantum}
cryptography, re-gaining some of what we have lost. The (im)possibility results for MiniQCrypt are nuanced, and provide
something between the intact conventional public key cryptography in Cryptomania, and the total absence of it in
classical Minicrypt.

In the discourse on quantum computing and its application to cryptography, it is important to be mindful of which
security notion the authors of some source, or the implementors of some device base their work on. Especially in
academic work, Pessiland assumptions are often implicitly made\cite{
    diamantiPracticalChallengesQuantum2016,
    kwekChipbasedQuantumKey2021,
    mehicQuantumKeyDistribution2021,
    loSecureQuantumKey2014,
}. Here, the speedup provided by Grover's algorithm is considered to make symmetric primitives like hash functions or
symmetric ciphers unusable, leaving only information-theoretically secure cryptographic schemes such as
one time pads available. In this framework, secret key rate becomes paramount because it is assumed that QKD keys will
be used with an information-theoretically secure encryption scheme, requiring an infinite, high-bitrate secret key
stream.
\todo{introduce notions of asymmetric/symmetric ciphers, OTPs before}

While in academic sources Pessiland assumptions are common, commercial systems usually are based on Minicrypt
assumptions. That is, commercial systems propose QKD as an alternative to classical asymmetric cryptography for
cryptographic key exchange, but then continue to use classical symmetric cryptography for purposes such as key
derivation and secret-key encryption. Using a computationally secure key derivation function such as Argon 2, a small,
fixed amount of precious QKD secret key bits can be expanded into a key of almost unbounded length\footnote{Key
derivation functions have limited output size}. Similarly, a
computationally secure symmetric cipher such as AES can be used to encrypt almost arbitrary amounts of data using a
single, short key\footnote{
    We write that the amount of data that can be encrypted with a computationally secure block cipher is only
    \emph{almost} unbounded because the cipher operates on blocks of a fixed, short size and depending on the cipher
    mode, in most applications, collisions of two such blocks enable stochastic \emph{Birthday
    Attacks}\cite{giraultGeneralizedBirthdayAttack1988}. Usually, for a primitive of block size $n\;\unit{\bit}$, an
    amount of $2^\frac{n}{2}$ extracted blocks is used as an upper bound for safe usage. For a cipher using the
    currently common block size of \qty{128}{\bit}, this bound lies at \qty{256}{\exa\byte} of
    data\cite{bhargavanPracticalSecurity64bit2016,}.
}.

\subsection{The Practical Security Implications of Quantum Computing}
\label{qc-practical-implications}

Given that as of yet, noone has claimed to have a quantum computer powerful enough to pose a threat to current
cryptographic protocols\cite{roettelerQuantumResourceEstimates2017}, one may ask the fair question why the possible
future development of such a machine would be consequential for today's cryptographic practice. The answer to this
question lies in \emph{Store-Now-Decrypt-Later} attacks. In such attacks, the attacker records all data transmitted
between a cryptographic protocol's parties. The security of any key exchange protocol rests on a computational hardness
assumption about some particular problem. When this assumption falls, for example because of a powerful quantum computer
becoming available, the attacker can then retroactively break the security of those stored protocol instances and
decrypt all traffic.

Modern cryptographic protocols such as TLS or the Signal messenger's key ratchet are designed with facilities to provide
some degree of protection against key compromise called \emph{(Perfect) Forward Secrecy}. Forward Secrecy means that a
compromise of keys at one protocol step will not break the secrecy of past protocol steps. Forward Secrecy is achieved
by repeatedly mixing fresh key material called \emph{Ephemeral Keys} into the protocol's secret state. For a
post-quantum attacker, this implies that to decrypt a run of a forward-secret cryptographic protocol, the quantum
algorithm breaking the protocol's computational assumption must be run a number of times, but this results only in a
linear increase of both protocol and attack complexity, which turns out to no advantage for the defender.

Store-Now-Decrypt-Later attacks are considered a serious threat today based on the stark discrepancy between the
capacity of today's inexpensive storage media, and the comparatively tiny bandwidth of cryptographic protocols in
applications such as End-To-End Encrypted (E2EE) text messaging. A single hard drive can conceivably store years of a
person's encrypted digital communications.

There has been ongoing work on quantum secure cryptographic algorithms, and standardization of several such algorithms
is progressing. However, in the time frame of cryptosystems, these algorithms are still rather young and the recent
discovery of a catastrophic key recovery attack against the Supersingular Isogeny Diffie-Hellman protocol
(SIDH)\cite{castryckEfficientKeyRecovery2023} illustrates the risk in the use of immature cryptographic primitives. Thus,
recommendations on the concrete steps that should be taken today to mitigate Store-Now-Decrypt-Later attacks vary. For
instance, under its threat model as laid out in \textcite{schmiegGoogleThreatModel2024}, Google recommends a list of
quantum secure counterparts to classically secure cryptographic algorithms, but recognizes the relative immaturity of
these quantum secure algorithms and consequently recommends \emph{Hybrid Deployment}, where a young, quantum secure
algorithm is paired with a mature classically secure algorithm such that \emph{both} algorithms would have to be broken
to compromise the composite protocol's security. Given that quantum secure public key cryptography tends to have both a
much larger key and/or ciphertext size and worse performance compared to state-of-the-art Elliptic Curve-based key
exchange or signature algorithms, pairing it with a classically secure alternative incurs only a negligible overhead in
key storage, network communication and computation costs.

\todo{research some more policies.}

\section{The Physics of Quantum Computing}
\todoplaceholder{missing}

\section{Quantum Key Distribution}

As we discussed in Section \ref{qc_comp_assum}, Quantum Computers promise novel attacks on many contemporary
cryptographic systems. At the same time, quantum technology also promises new cryptographic primitives that support
security guarantees beyond what can be realized with the best classical computers. The core of this nascent field of
Quantum Cryptography is a set of methods that are collectively called Quantum Key Distribution.

Informally speaking, a Quantum Key Distribution system is a system that distributes a secret key between two\footnote{
    Although the key distribution problem can conceptually be framed for any number $n\ge 2$ of parties, practical
    treatment is almost always limited to the two-party case. In case of QKD, problem instances for $n > 2$ parties can
    trivially be reduced to $(n^2 - n)/2$ invocations of the two-party protocol, combined with any
    information-theoretically secure secret sharing scheme.
} parties such that after a successful execution of the protocol, each of the two parties holds a copy of a randomly
generated secret key, and the probability that an attacker was able to extract some portion of the key during the
protocol's execution can be bounded to some negligible $\epsilon$ by each of the parties.

Quantum Key Distribution provides a similar service to cryptographic key exchange protocols such as the classic
Diffie-Hellman key exchange provide. The core difference between QKD and cryptographic key exchange protocols is that
QKD provides information-theoretic security based on the No-Cloning Theorem, where cryptographic protocols provide only
computational security based on the computational hardness assumption underlying some public key cryptosystem.

QKD is attractive in that it gives practically useful security guarantees without relying on any computational hardness
assumptions. This way, QKD would remain secure even in a scenario where a hybrid deployment of a classically secure but
mature algorithm paired with a quantum secure but young algorithm as discussed in Section
\ref{qc-practical-implications} poses too much of a risk---a scenario where both large quantum computers arrive and a
flaw in the quantum secure algorithm is found. Note that here, because we assume we have large quantum computers, the
possibility of a flaw in the quantum secure algorithm extends beyond mathematical flaws leading to practical attacks
with classical computers, and includes novel quantum algorithms.

\subsection{Security assumptions in QKD}

While QKD protocols provide information-theoretic security, part of these protocols is always an authenticated channel
that is used by the protocol's parties to exchange information necessary to align both parties' quantum measurements so
that they can reconstruct the same secret key bit stream\cite{loSecureQuantumKey2014}. In the security model of QKD,
this authenticated channel does some heavy lifting. While the QKD protocol provides key exchange--an asymmetric
primitive--based on this authenticated channel--which in its most simple implementation requires only symmetric
primitives, an implementation of QKD using symmetric primitives such as HMAC or CMAC for the authenticated channel would
not achieve information-theoretic security. To achieve information-theoretic security, the authenticated channel itself
must use an information-theoretically secure authentication method. The issue with that is that
information-theoretically secure authentication methods are (provably)\todo{citation on ``provably''} rather inefficient
in their key use. While symmetric MACs can use a single, short key for a very long time, information-theoretically
secure MACs need a continuous stream of fresh key bits.

In QKD, the authenticated channel can be bootstrapped by taking these MAC key bits from the QKD channel itself. The
disadvantage of doing that is that it consumes a fraction of the system's precious secure key rate. As a consequence, at
this point there is ongoing research\todo{citations on ongoing research} on both systems based on symmetric MACs and
systems using information-theoretically secure MACs, with commercial systems often choosing the
latter\cite{bibakQuantumKeyDistribution2021} owing to the low secret key rates that are the state of the art.
\todo{Finish this section}

\subsection{The Technical Implementation of QKD}

On the technical level, QKD must be distinguished from general Quantum Computing. While QKD systems employ the
No-Cloning Theorem and sometimes quantum entanglement in their operation, the scope of their quantum operations is very
limited. QKD systems always operate on photons, while general quantum computers use a variety of physical
implementations for their qubits that include photons and squeezed light, but extend over atom nuclei, trapped ions,
various aspects of currents in superconducters as well as phonons\cite{berriosHighFidelityQuantum2012}.

\todoplaceholder{Add concrete description of at least one QKD protocol (BB84?)}

\subsection{Practical Challenges}

The central challenge in general quantum computers is extending the lifetime of the quantum state encoding a qubit.
Quantum states are extremely sensitive to disturbances, and despite the best efforts to shield them against external
influence, their lifetime is still inconveniently short compared to the timescales required for quantum computation,
resulting in significant amounts of noise in the output of quantum algorithms run on contemporary quantum
computers\cite{yetisInvestigationNoiseEffects2021}. Quantum Key Distribution systems use photons and only perform a
handful of operations on each photonic state between generation and measurement, with the vast majority of the state's
lifetime spent in transit between the two endpoints of the QKD protocol.

While QKD systems are easy to build and operationally robust compared to general quantum computers, at their core they
still exchange information through quantum states that physically need to transit the distance from one endpoint to the
other. For classical computer networks, bridging distances of hundreds or thousands of kilometers is no big challenge.
Using appropriate high-power transceivers, a single, \emph{unrepeatered} span of an optical link can bridge hundreds of
kilometers while simultaneously achieving data rates of several terabits per second. Longer ranges are regularly
achieved through the use of (analog!) optical amplifiers, with recent \emph{repeatered} systems approaching the petabit
per second boundary
\footnote{
    cf.\ this encyclopedic entry \cite{JUNOSubmarineNetworks}, press releases by participating companies
    \cite{NECBuildNew, NewCompanyBuilds}.
}. These classical optical systems operate at hundreds of milliwatts of optical power, a limit resulting from nonlinear
effects in the optical fibers used, power limitations of optical amplifiers, and limitations in power delivery to these
amplifiers.

In contrast, QKD systems operate on signals that are weaker by several orders of magnitude. While classical optical
signals use millions of photons per bit, the quantum states at the core of QKD systems must necessarily be ``weak''. A
single quantum state in the fiber on average should consist of approximately a single photon. If the system's quantum
states consisted of more than one photon carrying the same information, this would enable a \emph{Photon Number
Splitting Attack}, in which an attacker extracts one of the state's photons for later analysis, and forwards the
remaining photons to the receiver\cite{loSecureQuantumKey2014}. The attacker can then later measure the captured photons
to extract the same information that the receiver measured. In practical QKD setups, attenuated pulsed lasers are often
used, as there are no practical single-photon sources. The laser and its attenuator are tuned such that the average
photon count of a pulse is in the order of $0.1$ \cite{loSecureQuantumKey2014}. For such setups, mitigations exist that
prevent photon number splitting attacks\cite{wangBeatingPhotonNumberSplittingAttack2005}. However, while these
mitigations patch this security weakness for weak, attenuated pulsed lasers, they still do not allow for higher transmit
power.

The practical implication of this is that the optical brightness of a QKD system is directly proportional to the rate at
which the system can prepare, and later measure the individual quantum states. The primary limitation is the speed and
recovery time of the single-photon detector. In contrast to e.g.\ a simple photodiode that (mostly) linearly converts
incident photons into electron flow, SPDs are designed to provide a large intrinsic gain. This improves their bandwidth
as each photon's pulse must charge the detector's own parasitic capacitance as well as that of any wiring between it and
the frontend preamplifier, but in many detector designs this intrinsic amplification process is also the origin of a
long recovery time that limit's the detector's possible repetition rate. With today's electronics, repetition rates up
to a few \unit{\GHz} are feasible\cite{grunenfelderFastSinglephotonDetectors2023}. Alas, the brightness limit interacts
poorly with the reality of optical communication, especially through fibers. Even modern, high-quality fiber-optic
cables have attenuation in the order of \qty{0.2}{\dB\per\km}\cite{chesnoyUnderseaFiberCommunication2015}, which
corresponds to roughly half of the signal being lost every \qty{15}{\km}. In classical optical networks, this can be
compensated by increasing transmit power--i.e. packing more photons into each bit--or by optically amplifying the signal
partway through the fiber. cIn QKD systems however, the signal's quantum states cannot be amplified both out of a
concern of photon number splitting attacks and because of decoherence\footnote{
    Note that this impossibility is not a consequence of the No-Cloning Theorem. The No-Cloning Theorem only asserts
    that it is impossible to create a second, \emph{independent} copy of an arbitrary quantum state, which can then
    independently be measured without disturbing the original state. Despite this, a hypothetical ``quantum amplifier''
    could increase the quantum state's photon number, adding entangled photons that share the original quantum state.
    Alas, doing this would not gain us much in a QKD system because an interaction of any of the quantum state's photons
    with the fiber---that is, the same loss as before---would disturb the entire entangled state.
}, and thus the system's bit rate decreases exponentially with distance due to attenuation. Some QKD systems can reach
ranges of several hundred kilometers, but the resulting payload data rate---usually called \emph{secret key rate}---of
these long distance systems is measured in kilobits per second. An interesting observation from theoretical work on
quantum key distribution algorithms is that not only is this exponential rate decay a fundamental limit for a given QKD
implementation, but it is even possible to determine a protocol-independent upper bound for a noiseless, lossy optical
channel's secret key rate. This upper bound shows the same exponentail decay and, notably, is independent of the optical
power, which is directly proportional to the repetition rate of the QKD protocol's measurements. Modulo some small,
constant factor, this upper bound cannot be circumvented with any amount of protocol engineering, or source or detector
improvements\cite{takeokaFundamentalRatelossTradeoff2014}.

\subsection{Loss in optical fibers}

When transmitted over a fiber, there are multiple effects that degrade the quantum-optical signal of a QKD system, which
are collectively referred to as \emph{loss}. We can coarsely classify these degrading effects into two categories:
\emph{decoherence}, and \emph{attenuation}. Decoherence effects result in the quantum state being changed in transit,
which depending on the QKD implementation may mean destroying information contained within the state such as by
disturbing the pulse's polarization, or destruction of entanglement between the in-flight state and another local state.
In contrast, attenuation means the quantum state is not ever leaving the channel.

In practice, attenuation is the primary factor limiting the length of an individual fiber run in QKD. Even modern,
ultra-low loss optical fiber has an attenuation in the order of \qty{0.15}{\decibel\per\kilo\meter}, resulting in a loss
of half the signal's power, equivalent to half of all QKD pulses, in just \qty{20}{\kilo\meter}. For longer reaches,
these losses ar multiplicative, so after only \qty{200}{\kilo\meter} only one in a thousand single-photon pulses entering
the fiber will exit it at the other end \cite{chesnoyUnderseaFiberCommunication2015}.

Decoherence effects are less relevant for the distance limitation, and mostly limit which fiber-optic technologies can be
utilized in the first place. Due to decoherence, QKD systems usually use Single-Mode (SM) fiber over Multi-Mode (MM)
fiber\cite{amitonovaQuantumKeyEstablishment2020}, and decoherence makes it more difficult to utilize Wavelength Division
Multiplexing (xWDM) to send multiple either quantum or classical optical signals through a single fiber.\todo{is this
right?}

Attenuation in optical fibers has a number of origins. The main factor is scattering of photons on the fiber core, with
absorbtion due to interactions between photons and the fiber core's molecular structure or embedded contaminants only
playing a minor role. The primary component of scattering is fluctuations in the fiber core's molecular structure, with
scattering on phonons (Brillouin scattering) or photons (Raman scattering) only adding a samll amount of
loss\cite{wandelAttenuationSilicabasedOptical2006}.

Like attenuation, decoherence can also result from a number of different mechanisms. Two optically \emph{linear}
mechanisms, i.e.\ ones that do not depend on incident signal power, are chromatic dispersion and polarization mode
dispersion (PMD). PMD disturbs the signal's polarization. PMD strongly depends on wavelength and is highly sensitive to
environmental factors such as temperature or vibration \cite{brodskyPolarizationModeDispersion2006}. QKD systems
frequently use polarization-based encodings, which are sensitive to PMD. PMD is usually mitigated by continuously
measuring the fiber's end-to-end PMD, and adjusting a polarization controller placed
in-line\cite{wangLongdistanceCopropagationQuantum2017, ImpactPolarizationMode,
agnesiAllfiberSelfcompensatingPolarization2019} with the fiber to cancel out the fiber's PMD.

Chromatic dispersion arises from the fiber's materials' refractive index not being perfectly constant across
the spectral bandwidth of the optical signal, leading some frequency components of the signal to traverse the fiber
faster than others, resulting in pulses being spread out as they continue along the fiber. Chromatic dispersion is a
concern in some long-distance QKD systems that need to operate at a timing precision down to a few dozen picoseconds,
but like PMD it can be compensated at the endpoint \cite{neumannExperimentallyOptimizingQKD2021,
kiselevAnalysisChromaticDispersion2020}.

Besided linear Brillouin and Raman Scattering, nonlinear effects such as the AC Kerr Effect, Stimulated Raman Scattering
as well as Stimulated Brillouin Scattering can produce intermodulation and crosstalk when a quantum optical signal is
sent through the same fiber as another, much brighter classical optical signal. These nonlinear effects are relevant for
QKD systems that either send a reference clock through the same fiber as the QKD pulses, or that aim for coexistence
between QKD pulses and classical optical networking on the same fiber, for instance in an in xWDM
setup\cite{choiQuantumKeyDistribution2010, grunenfelderLimitsMultiplexingQuantum2021}.

In the AC Kerr effect, a strong optical signal influences the refractive index of the fiber core, which modulates other
signals propagating through the same fiber. Stimualated Brillouin Scattering arises when a high-power incident signal
causes the emission of phonons inside the fiber core, which then act as a source of Brillouin scattering. Stimulated
Raman Scattering is a similar effect based on Raman scattering\cite{chesnoyUnderseaFiberCommunication2015}. When a fiber
is shared between weak QKD and bright classical signals, both Brillouin and Raman scattering introduce noise in the QKD
channel as photons from the classical signal change their wavelength, and might end up inside the QKD channel's
bandwidth\cite{choiQuantumKeyDistribution2010}.

\todo{Some detail on CV-QKD}

\subsection{Relaying}
\todo{(one?) term of the art seems to be "repeater"}

We cannot use conventional optical amplifiers to extend the range of a single continuous QKD link lest we destroy the
signal or we might enable attacks. What remains as ways to extend the range of a QKD link are \emph{relaying} methods,
where one QKD link is terminated at a relay station partway to its destination, and another is started, with the relay
proxying information between the two. We can separate relay implementations into two broad categories.

\begin{description}
    \item[Classical relays] encompass the trivial implementation of a relay, where the QKD link is formed by simply
        stitching two QKD links together by connecting one link's receiver to the other link's transmitter. The key
        characteristic of classical relays is that inside the relay, the link's cryptographic payload information is
        handled in its classical plaintext form. Classical relays are practically feasible, but because they must handle
        the payload in plaintext form, they are security-critical.

    \item[Quantum relays] are relays that forward the QKD payload information from one link to the other in the quantum
        realm, without translating it to classical information and back. QKD relays are currently not practically
        feasible, but if they become available in the future, they would allow range extension without compromising the
        QKD link's security as the same tamper-detecting properties that the QKD links provide can be extended to cover
        the quantum forwarding process inside the relay.
\end{description}

For practical purposes, classical relays are the only relevant option. A long-range QKD system employing classical
relays would be able to cover arbitrary distances, trading off reliance upon physical security of the trusted relay
stations. Academic work on QKD recognizes this limitation, but few proposals to its solution have been put forth.

\subsection{Range extension in Measurement Device Independent (MDI)-QKD}

One technology closest to a solution on the trusted relay issue is Measurement Device Independent (MDI)-QKD. Broadly
speaking, in an MDI-QKD system two QKD endpoints are connected through exactly one relay (or router). The key idea of
MDI-QKD is to move all trusted components of the protocol out of this central relay, and into the trusted nodes at both
ends of the link. Instead of directly measuring the photons sent by both endpoints, the relay node has them interfere
and measures the result of this interference. This measurement result does not allow the relay to draw any conclusions
on the individual qubits that the endpoints exchange, but when the relay communicates these measurements to the
endpoints, the endpoints can reconstruct their shared secret key bits. Although in MDI-QKD the relay node still performs
quantum measurements and participates in the overall QKD protocol, the protocol guarantees that even a malicious relay
cannot learn anything about the exchanged keys from its limited vantage point.

MDI-QKD effectively doubles the range of a QKD system. Unfortunately, the approach from MDI-QKD cannot be adapted to
multiple chained relays, and thus it is mostly interesting for hub and spoke-style quantum network topologies. In a
relay-assisted long-range QKD system, MDI-QKD could only be used to eliminate trust in half of the relays, which in the
grand scheme of things does not reduce attack surface by much.

\todo{Mention entanglement swapping range extension}

\section{Quantum Networking}

So far we have focused on the range limitation of a single QKD link with classical relays as the only practical solution
at this point in time. Quantum Networks naturally follow from a relay-assisted QKD link, if we consider a type of
``relay'' that is connected to more than two links. Just like switches and routers can be meshed to construct complex
topologies in classical wide-area networks (WANs), such multi-fanout relays, or \emph{routers} can be used to provide
QKD services over complex network topologies.

There exists a large corpus of academic research on the theory of such large-scale QKD networks ranging from the
technical implementation of management protocols to specialized QKD systems for QKD networks that improve on standard
two-party QKD in areas such as complexity or performance. \todo{lots of citations here}
In the past decades, a number of proof-of-concept QKD networks have been put into practice. None of these systems
provide any practical utility yet, and their raison d'être lies in the political realm more than it arises out of
technical necessity considering that any of today's city-scale demonstrations can easily be simulated more compactly in
a lab using a few spools of fiber as a near-perfect stand-in for long-range fiber links.

Many of the technical challenges in the deployment of QKD networks coincide with similar technical challenges in
classical packet-switched networks. An unique challenge to QKD networks is how their routing problem is different to the
one in classical computer networks. In a classical network, each link has a known, fixed capacity. A router decides
which packet to send through which link, and when the rate of incoming packets momentarily exceeds the capacity of the
outgoing links, packets must either be dropped, or put into a growing queue. QKD networks are different in that
information is not exchanged through the network, but instead the network \emph{generates} information in the form of
secret key material. The measurement of individual pulses that underly key generation conform to a stochastic process,
but amortized across the large time spans required for the subsequent selection and privacy amplification steps that
converts these raw measurements into usable secret key bits, key generation rate is constant. Each node of a QKD network
thus accumulates secret key bits for each of its links, storing them for later use. The routing problem in this scenario
revolves around managing the levels of these key stores to avoid depletion.

\section{Securing QKD Networks with Inertial HSMs}

As we discussed above, when it comes down to practical, end-to-end security properties, Quantum Key Distribution
removes trust in the hardness of particular mathematical problems (good!), but increases trust in the physical
integrity of the transceivers of the QKD link (bad!). In scenarios where the communicating parties are all located
within physical proximity---in QKD, meaning within at most a few hundred kilometers from each other depending on secret
key rate requirements---this added trust is of no consequence because the communcating parties' hardware must be trusted
in either QKD-assisted or purely classical setups. However, this trust requirement becomes a burden as soon as at least
one party is too far away or when higher secret key rates are required, as now physically trusted relays become necessary.

Extrapolating to practical deployments, we can make two predictions. First, as QKD only solves key distribution, but the
actual data transfer still happens through normal off-the-shelf telecommunications components in QKD networks, there is
no reason for a practical QKD setup to \emph{not} also use classical cryptography as an additional layer for defense in
depth,
\todo{citation on defense in depth, and on this hybrid scenario}
meaning the QKD setup will at worst degrade to the same security a purely classical system would provide, never less.

The second prediction we can make is that any practical QKD network will have to use trusted relays to bridge large
distances. While in certain specialized applications such as the proposed financial QKD network in Switzerland
\todo{citation on swiss deployment} smaller, isolated networks are conceivable, in every telecommunication system from
the telegraph through the telephone system and up to the internet it has been shown conclusively that considering
utility, a global, interconnected network is greater than the sum of its parts\footnote{In fact, history repeats, and
the enthusiasm that Quantum Key Distribution networks have kindled parallels the one that the first trans-atlantic
telegraph cables brought forth as described by \textcite{mullerWiringWorldSocial2016}. Both parallel not just in the
extensive promises attributed to their respective technologies, but also in the facade of technological determinism that
in both cases hides a number of social and political motivations.}\cite{mullerWiringWorldSocial2016}. \todo{at least one
more citation on historic networks}

In this section, we will outline a solution that provides practical, end-to-end security in large-scale QKD networks by
delegating the hardware trust issue of QKD relays to Inertial Hardware Security Modules. The primary design challenges
we will address are the systems' overall envelope design, optical passthroughs, and matching the cryptographic
assumptions behind the IHSM's heartbeat and alarm subsystem to those of the QKD application.

\subsection{The anatomy of a QKD node}

With the exception of special cases such as the middle node in a MDI-QKD system, a general QKD relay contains the same
components that the endpoint of a QKD connection uses. Only in a QKD relay, two transceivers are connected back-to-back
to one another. QKD provides physical security for the photons traversing the fiber that forms the system's channel, and
the security envelope of the system begins where this fiber is terminated in the power splitters, single-photon
detectors, lasers, and interferometers of the QKD transmitter and receiver. To process the raw measurements of the QKD
system into a usable stream of secret key bits, in addition to these components implementing the physics of the QKD
system, a classical computer is needed. On top of the remote monitoring and management tasks that any piece of
networking equipment is expected to perform nowadays, this computer is tasked with the information reconciliation and
privacy amplification that form the information-theoretic part of the QKD system. Since this computer must necessarily
handle secret key bits in their plain text form, it, too, must be inside the relay node's physical protection envelope.

\subsection{Physical requirements of QKD transceivers}

Putting a QKD relay node and associated machinery inside of an IHSM, we first need to answer two key questions. First,
\emph{will it fit?}, and second, \emph{Can we hook it up?}. In the following paragraphs, we will go through several
aspects of these general questions one by one.

\paragraph{Physical dimensions.}
At this point, a number of commercial systems promising QKD exist. Common QKD protocols do not require any particularly
large or power-hungry components, and so commercial systems have generally adopted the 19 Inch rackmount enclosure
standard that is common to modern telecommunications equipment, with a width of $\approx\qty{50}{\centi\meter}$, a
height between $\approx\qtyrange{4}{30}{\centi\meter}$ and a depth below $\approx\qty{100}{\centi\meter}$.\todo{Re-check
these numbers shortly before submission} While something of this size would be infeasible to protect with the security
mesh of a traditional hardware security module, placed vertically, even without modifications any of these systems are
well within an envelope that can be protected with a single IHSM cage.

\paragraph{Power supply.}
QKD systems do not contain any particularly power-hungry components. Unlike quantum computers, most of the signal path
is optical, and as such can be implemented with room-temperature fiber-optic components. Only the single-photon
detectors may require cooling in some systems, but unlike something like an ion trap quantum computer's processor,
energy-intensive deep cryogenic cooling is not necessary. Most manufacturers don't quote the power requirements of their
systems, but we were able to find that IDQuantique specifies their QKD systems to be able to run off a single
\qty{300}{\watt} power supply\cite{ClavisXGQKD2024}. In an inertial HSM, power up to several \unit{\kilo\watt} can
easily be transferred to the payload with through-axis cables.

\paragraph{Cooling.}
While the few hundred Watt of power that QKD systems require could easily be transported through the mesh of a a
traditional HSM as well, cooling that amount of thermal load purely by heat conduction through centimeters of epoxy
resin would make implementation infeasible in traditional HSM. In an IHSM on the other hand, up to several
\unit{\kilo\watt} can easily be dissipated through forced-air cooling since the rotating security mesh can have an
arbitrary amount of longitudinal openings.

\paragraph{Data and signals.}
A QKD transceiver has a number of ports in addition the port for the fiber optic quantum channel. Depending on the
system, one or more additional optical links may be necessary for clock distribution, allowing both endpoints to tune
their lasers into precise alignment. QKD protocols require a classical link used for information reconciliation, which
along with the key stream output and management links requires one or more classical network ports.

In a QKD relay node, the key stream never leaves the security envelope. The management and information reconciliation
links can be combined into a single, classical network link, requiring a single fiber when using a standard wavelength
division multiplexing transceiver. The QKD link's clock channel and the quantum channel require a dedicated fiber each,
adding up to a total of five fibers for a uni-directional QKD relay, or nine fibers for a bidirectional one. Since fiber
pigtails have an outer diameter of usually about \qty{1}{\milli\meter}, this amount of fibers can be fed through an
IHSM's axis of rotation. The mechanical challenge in such a multi-fiber signal and data feedthrough is to observe the
fiber's minimum bending radius, which for common fibers is usually in the range of
\qtyrange{5}{15}{\milli\meter}\cite{fs1M12FSC,ProductPageFiber,CorningSMF28Ultra2024}.

Concluding the above paragraphs, a QKD node is not a particularly challenging payload for an IHSM. The most problematic
requirement is feeding through a number of fibers for its various input and output signals, but fundamentally it is no
different from any server or other piece of IT equipment. In the following section, we will present a design that
provides a combined power and multi-fiber passthrough that is sufficient for QKD applications before concluding with an
analysis of post-quantum heartbeat signal security.

\subsection{Multi-fiber passthrough with active secondary mesh}

The primary weak spot of a simple IHSM is its axis of rotation. While the stationary axis allows for wired data and
power connections to penetrate the mesh, it also provides an easy target for an attacker who wants to insert some sort
of physical probe into the IHSM's security envelope. While to a certain extent this attack vector can be made more
difficult though simple construction techniques such as making the shaft as thin as possible, and getting the mesh as
close to it as possible, as well as using a solid steel shaft on the motor end of the mesh, the level of security that
these mitigations provide is much below that of the remainder of the mesh. Thus, a better solution is needed.

Previously, in Chapter \todoplaceholder{provide link to mesh protection overview from OG IHSM paper} we have alluded to
several \emph{shielding} methods that use a independently rotating secondary mesh on the inside of the primary mesh,
located right next to the primary mesh's axis opening. In this section, we will go into some more detail on four
variations of this solution. In order of increasing complexity, these variations are a simple disc cover, coaxial
labyrinth meshes, offset labyrinth meshes, and interlocking gear meshes. We will demonstrate a functional prototype of
the simple disc cover, present a design and mechanical prototypes of the offset labyrinth meshes, and provide details on
the design of a interlocking gear mesh.

\subsection{Simple disc cover}

\todo{Update these graphics with final color scheme, and update caption text here}

\begin{figure}[h!]
    \centering
    \includegraphics[width=\textwidth,page=1]{shaft_countermeasures_b.pdf}
    \caption[Coaxial disc mesh schema]{\draftgraphics Coaxial disc mesh schema, cross-section and top-down views. The
    outer mesh is shown in red, and the inner mesh in blue. The dashed line indicates the two meshes' shared axis of
    rotation. The gray areas indicate the shape of the volume that remains undisturbed by the mesh, and that is
    available for structural support and cable routing.}
    \label{qkd_fig_disc_mesh}
\end{figure}

In Chapter \todoplaceholder{Provide link to single-board IHSM chapter here}, we have shown how an IHSM that has been
shrunk to a single, disc-shaped PCB is still useful because we can delegate key management functionality to the mesh
monitoring circuit's microcontroller---or a separate processor sitting next to it---on the rotating mesh PCB, yielding a
solution close in both its cryptographic capabilities and its security level to commercial traditional HSMs, and
exceeding those of a smartcard. In the following paragraphs, we will show how we can deploy the same single-board IHSM
(SB-IHSM) as a mitigation for through-axis attacks, exploiting its mechanical shape and its simple, low-cost
implementation.

By placing an adapted single-board IHSM close to the primary mesh's axis opening as shown in Figure\
\ref{qkd_fig_disc_mesh}, an attacker is forced to either first circumvent or at least dislodge the single-board IHSM
through the primary mesh's axis opening without disturbing either mesh to gain direct access to the payload behind it,
or to conduct their attack through the keyhole-sized opening in the primary mesh while bending their tool by
approximately \qty{90}{\degree} at least twice, once to avoid the SB-IHSM mesh, and once more to re-orient the tool
towards the payload. The distance between the inside of the primary mesh and the SB-IHSM is limited by the tolerance in
mechanical alignment between the two axes of rotation, by the space necessary for a sufficiently stable mount of the
payload cage to the hollow shaft, and by the minimum bend radius of the power and data wiring that needs to pass through
the shaft. In QKD applications, the fibers' minimum bend radius is the largest contributing factor. Power and electrical
data signals can be supplied through flexible flat cables that can be bent in sharp corners without issue. Optical
fibers on the other hand are limited in their minimum bend radius, as their optical loss rises sharply with decreasing
bend radius\footnote{Note that the issue here is not that the glass core of the fiber would degrade or break, as one
might intuitively assume. Being only a few dozen micrometers in diameter, an optical fiber's core is remarkably
flexible. Instead, the issue is that both multimode as well as singlemode fibers are optical waveguides. Bending them
distorts the electromagnetic field inside the waveguide, and allows some small portion of it to escape from the fiber's
core, leading to loss in the form of both attenuation and dispersion\cite{schermerImprovedBendLoss2007}.}. With QKD
being especially sensitive to even small amounts of loss, care has to be taken to maximize the bend radius of the fiber
optic connections. A common specification of minimum bend radius in telecom singlemode fibers taking into account not
just optical loss but also the mechanical stability of the fiber's polymer coating is $10\times$ the coated fiber's
diameter\cite{fs1M12FSC,ProductPageFiber,CorningSMF28Ultra2024}, which equates to \qty{9}{\milli\meter} for common
\qty{0.9}{\milli\meter} fiber pigtails, corresponding to approximately \qty{1}{\decibel} of loss in the
\qty{1550}{\nano\meter} band\cite{schermerImprovedBendLoss2007}. Based on these specifications and on a conservative
estimate of \qty{2.5}{\milli\meter} for the vertical mesh clearance, we arrive at a minimum inter-mesh spacing of
approximately \qty{11}{\milli\meter} when using minimal overlap between tab heights. 

\begin{figure}
    \centering
    \subcaptionbox[Helical transition of single fiber]{Single fiber}{\includegraphics[width=.45\textwidth]{\scaledgraphics{helix_transition.png}}}
    \hfill
    \subcaptionbox[Helical transition of fiber bundle]{Fiber bundle}{\includegraphics[width=.45\textwidth]{\scaledgraphics{helix_bundle.png}}}
    \caption[Helically coiling fibers inside the axis tube]{
        The necessary mesh spacing can be reduced by coiling the fibers inside of the axis tube. The coiled fibers enter
        the inter-mesh space at an angle equal to the helix lead angle, which reduces the amount of space necessary to
        complete the transition to horizontal along a circular arc. In this example, a \qty{6}{\milli\meter} outer
        diameter tube with a \qty{0.5}{\milli\meter} wall thickness is shown with 6 fibers with \qty{0.9}{\milli\meter}
        outer diameter coiled to a constant bend radius of \qty{9}{\milli\meter}. The lead angle of the resulting helix
        is \qty{61.5}{\degree}, and past the tube exit, only \qty{5.16}{\milli\meter} of inter-mesh space are necessary.
        \figureattrib{helix_transition.png}}
    \label{qkd_fig_fiber_helix}
\end{figure}

\todoplaceholder{Finish this part. Use the rev 1 SB-IHSM to build a practical prototype.}

\subsection{Coaxial labyrinth meshes}

\begin{figure}[h!]
    \centering
    \includegraphics[width=\textwidth,page=4]{shaft_countermeasures_b.pdf}
    \caption[Coaxial labyrinth mesh schema]{\draftgraphics Coaxial labyrinth mesh schema, cross-section and top-down
    views.}
\end{figure}

In QKD applications, the simple disc cover design shown above has two main limitations. First, the distance between the
primary and secondary meshes' tab rings must be large enough to allow for the fibers' minimum bend radius, resulting in
more than \qty{10}{\milli\meter} of space available to an attacker. Second, the attacker only has to bend their tool in
a plane to reach the payload.

To increase the difficulty of inserting a long and flexible tool through the axis shield, \todo{Axis shield might be a
nice term. Unify terminology for axis/shaft, the shield, the names of the two meshes, and the tabs sticking up from the
meshes. Also what do we call the space in between? Terminology for the sides with offset meshes?} the shape of the
interface layer between the two meshes can be made more complex. Introducing small mesh \emph{tabs} that stick out
into the inter-mesh space from both meshes creates a labyrinth-like structure between the axis opening and the IHSM's
inside. Structural support and cables can easily pass this structure in a series of \qty{90}{\degree} bends, while
inserting a probe avoiding both meshes would not be feasible as the probe would have to perform a series of sharp
bends. The type of manipulator that would be necessary for the placement of a probe in this system is conceptually
similar to snake-like robots used in minimally invasive surgery, but state-of-the-art systems from this area are both
too thick and don't have enough joints to fit even simple labyrinth layouts\cite{
    suhDesignDiscreteBending2017,
    schmitzRollingTipFlexibleInstrument2019,
    kimAdvancementFlexibleRobot2022,
    hongDesignCompensationControl2020}.
For instance, if we assume \qty{3}{\milli\meter} material thickness on the radial bracket connecting the shaft with the
secondary mesh's mounting frame\todo{conceptual drawing here} along with \qty{10}{\milli\meter} of mesh tab overlap,
\qty{1.5}{\milli\meter} of clearance between radial bracket and each of the two meshes, and an inter-mesh spacing from
one tab ring to the next equal to the radial brackets' material thickness of \qty{4}{\milli\meter} plus the clearance
from bracket to mesh, we arrive at a meander \qty{6}{\milli\meter} in width completing four \qty{180}{\degree} turns
within less than \qty{40}{\milli\meter} of radial distance.

Researching the security of nuclear weapons, \textcite{bellovinPermissiveActionLinks} references a quote characterizing
the tamper security of a Permissive Action Link, a tamper-proof component designed to authorize the use of a nuclar
weapon through a code, as follows.
\todo{Get the actual book from ULB, and properly attribute this quote.}

\begin{quote}
    Bypassinag a PAL should be, as one weapons designer graphically put it, about as complex as performing a
    tonsillectomy while entering the patient from the wrong
    end. \cite{caldwell1989reducing,bellovinPermissiveActionLinks}
\end{quote}

With our discussion of surgical robots two paragraphs ago this quote is very on the nose, and it is probably fair to say
that we have made some progress to achieve this standard. While we are not quite there yet, we shall make it our goal to
achieve or even exceed this standard with our work in the following sections.

\begin{figure}
    \centering
    \includegraphics[width=.7\textwidth]{\scaledgraphics{wikimedia_Four_Corners_Bank_Vault_cropped.jpg}}
    \caption[Photo of a bank vault door]{\camerareadygraphics Photo of a bank vault door at the Four Corners building in
    Bowling Green, Ohio, USA. The interface between the door and its frame is stepped all around to discourage would-be
    intruders from inserting any sort of tool through the small gap around the closed door. In this instance, because
    the door's sill is stepped, too, a small ramp has been placed over the sill so that people going in and out of the
    open door don't stumble over the steps.\\
    \imgsource{Wikimedia Commons user Mbrickn}{2019}{CC-BY-SA}{https://commons.wikimedia.org/wiki/File:Four_Corners_Bank_Vault.jpg}
    }
    \label{qkd_fig_vault_door}
\end{figure}

While long and narrow tabs are desirable for mesh security as they limit the size and mobility of an attacker's probe,
in QKD application, the need for fiber optic passthrough is the limiting factor. The obvious solution of passing through
the fibers in a series of in-plane S-bends requires a coarse tab spacing due to the fibers' large minimum bend radius.
However, we can apply the approach we proposed above for the shaft entrance here, too, and thread the fibers between the
meshes by helically coiling them, increasing the fibers' bend radius to one half of the distance between both mesh
discs minus the fibers' diameter and clearances\todo{Formulas here and elsewhere, define variables}. When the resulting
useable part of the distance is larger than twice the bend radius, the minimum tab spacing is only limited by the
fiber's diameter and the stability of the star bracket. When the discs are placed closer, and a larger pitch is
necssary, the resulting pitch of the helix determines the minimum tab spacing.

Designing a labyrinth mesh for intrusion prevention is similar to the design of the shape of the jamb of a safe door
such as the one shown in Figure\ \ref{qkd_fig_vault_door}, or of a high end apartment door. In these, the objective is
to prevent would-be burglars from inserting opening tools through the space between the closed door and its jamb and
attacking the door's interior handle or locking mechanism, not unlike an IHSM's defense against electrical or
electromagnetic probes. The one difference between these doors and what we can do in IHSMs is that these doors are
limited to outwards-facing steps because they must be opened and closed. In IHSM labyrinth meshes, we can use both
outwards-facing and inwards-facing steps.

Concentric labyrinth meshes allow for a wide range of different configurations. The pitch from one mesh tab to the
next is the sum of the required width of the inter-mesh space and the safety margin needed betwween any cables or the
inter-mesh bracket and the tabs. When the mesh is constructed using rigid PCB tabs that are inserted as-is, without
bending them, and when all tabs have the same width and thickness, the radial width of the swept area decreases from tab
to tab going outwards as shown in Figure\ \ref{qkd_fig_mesh_ring_reduction}. A consequence of this is that when the
design target are constant width inter-mesh spaces, the tabs' pitch decreases going outwards.

\begin{figure}
    \centering
    \includegraphics[width=\textwidth]{mesh_ring_reduction.pdf}
    \caption[Coaxial labyrinth mesh tab swept area]{\draftgraphics Top-down view of a coaxial labyrinth mesh
    with three tabs, with the area swept by each tab highlighted. When rigid, planar tabs of a single width $w$ are
    used, the radial width of the swept areas decreases and approaches the tabs' thickness $t$ as their radius $r$
    increases.
    }
    \label{qkd_fig_mesh_ring_reduction}
\end{figure}

The safety margin required to avoid collisions between the meshes and the stator\todo{stator is a nice word for the
entire non-rotating part of the assembly. stator/star bracket?} can be kept low for the primary mesh because this mesh
has high-quality bearings on both ends, leading to good axis alignment. In contrast, for the secondary mesh considerable
margins have to be included if the mesh is driven by a cooling fan motor, as the bearings in such fans are not very
precise. With loose bearings, angular axis misalignment can lead to several millimeters of deflection in both the radial
and axial dimensions as illustrated in Figure\ \ref{qkd_fig_mesh_ring_bearing_tolerance}.

\begin{figure}
    \centering
    \includegraphics[width=\textwidth]{mesh_ring_bearing_tolerance.pdf}
    \caption[Coaxial labyrinth mesh axis alignment tolerance illustration]{\draftgraphics Illustration of the effect of
    angular misalignment of the axis of rotation caused by tolerances in motor bearings in a coaxial labyrinth mesh with
    two tabs. The area swept by each tab, and its increase due to misalignment are highlighted. The left illustration
    shows the ideal and misaligned meshes, and the right illustration superimposes the area increase from the left
    illustration on the ideally aligned mesh. This illustration is not to scale.}
    \label{qkd_fig_mesh_ring_bearing_tolerance}
\end{figure}

\subsection{Offset labyrinth meshes}

\begin{figure}[h!]
    \centering
    \includegraphics[width=\textwidth,page=2]{shaft_countermeasures_b.pdf}
    \caption[Offset labyrinth mesh schema]{\draftgraphics Offset labyrinth mesh schema, cross-section and top-down
    views. The two dashed lines indicate the two meshes' offset axes of rotation, shifted in $x$ direction in both
    views.}
    \label{qkd_fig_offset_lab_schema}
\end{figure}

Concentric labyrinth meshes improve upon simple disc meshes in security, but they have two remaining weaknesses. One is
that in a concentric labyrinth mesh, the part of the inner mesh at the axis is easily accessible through the opening in
the outer mesh. As the axis of rotation is the most vulnerable spot in a mesh because the tangential velocity of the
mesh is lowest close to the axis, tampering can be made more difficult by placing the axis of rotation of the inner mesh
not concentric with that of the outer mesh, but at a radial \emph{offset}.

A consequence of placing the axis of the inner mesh at an offset is that the inter-mesh rings formed by the tabs of the
two meshes now no longer form a set of concentric rings, but a set of nested non-concentric annulus shapes whose narrow
and wide sides alternate along the direction of the offset. We will show below how an optical fiber can still be wound
through this complex inter-mesh space without much trouble through a variation of the helical spiral trick from above to
avoid the annular rings' narrow sections. At the same time, the alternating narrow sections of the annular rings make it
more difficult to feed through the type of surgical robot we cited above, whose joints are designed for in-plane
operation for most of the manipulator, starting from the high-flexibility joints close to its end and down the neck. In
this section, we will show a design and a mechanical prototype of an offset labyrinth mesh design that improves on a
concentric labyrinth mesh on both the shielding of the secondary mesh axis and the feasibility of an attack with a
surgical robot without increasing mechanical complexity compared to a concentric design. In addition, we show a fiber
feedthrough that improves on the simple helical feedthrough we introduced above.

\begin{figure}
    \centering
    \includegraphics[width=\textwidth]{schema_wire.eps}
    \caption[Offset labyrinth mesh schema with fiber layout]{\figureattrib{schema_wire.svg}}
    \label{qkd_fig_offset_lab_fiber}
\end{figure}

Our offset labyrinth mesh design combines an offset of the secondary mesh's axis of rotation with the labyrinth mesh
approach from the previous section, creating wide and narrow inter-mesh spaces on alternating sides of the offset
direction as shown in in Figure\ \ref{qkd_fig_offset_lab_schema}. Structural support is provided using a CNC machined or
3D printed part, which also serves as a conduit for electrical connections from the shaft to the payload using Flexible
Flat Cable (FFC). While the FFC can easily conform to the offset labyrinth's sharp corners, an optical fiber can not.
Thus, instead of passing it straight through the labyrinth, the payload's fiber optic connections are passed through the
labyrinth in a three-dimensional spiral shape, avoiding the meshes while simultaneously maximizing the fibers' bend
radii.

To prove the mechanical viability of the offset labyrinth mesh concept, we created a mechanical prototype of one such
mesh. Figure\ \ref{qkd_fig_offset_lab_fiber} shows the dimensions of the meshes' tabs along with the resulting tab rings
and a 2D projection of our chosen fiber layout. The fiber is laid out in such a way that it crosses each tab ring at
opposite sides, and traverses the vertical distance in the larger part of the inter-mesh space. Figures\
\ref{qkd_fig_lab_mesh_exp_1} and \ref{qkd_fig_lab_mesh_exp_2} show an exploded view of our mechanical prototype from two
perspectives, and Figure\ \ref{qkd_fig_lab_mesh_section} shows a section view.

\begin{figure}
    \centering
    \includegraphics[width=\textwidth]{\scaledgraphics{render_exp_1.png}}
    \caption[Offset labyrinth mesh assmbly exploded render]{\figureattrib{render_exp_1.png}}
    \label{qkd_fig_lab_mesh_exp_1}
\end{figure}

\begin{figure}
    \centering
    \includegraphics[width=\textwidth]{\scaledgraphics{render_exp_2.png}}
    \caption[Offset labyrinth mesh assmbly exploded render]{\figureattrib{render_exp_2.png}}
    \label{qkd_fig_lab_mesh_exp_2}
\end{figure}

\begin{figure}
    \centering
    \includegraphics[width=\textwidth]{example-image-10x16.pdf}
    \caption[Offset labyrinth mesh assmbly exploded render, section view]{\draftgraphics\\
    Section view of the labyrinth mesh assembly}
    \label{qkd_fig_lab_mesh_section}
\end{figure}

\subsection{Interlocking gear meshes}

\begin{figure}[h!]
    \centering
    \includegraphics[width=\textwidth,page=3]{shaft_countermeasures_b.pdf}
    \caption[Offset gear labyrinth mesh schema]{\draftgraphics Offset gear labyrinth mesh schema, cross-section and
    top-down views. In this example, the axis is shifted by about twice the offset from the previous offset labyrinth
    mesh schema in Figure\ \ref{qkd_fig_offset_lab_schema}.}
\end{figure}

The offset labyrinth design already achieves a high level of security through its complex passthrough shape, but only
small offset distances are feasible since large offsets quickly lead to impractically large mesh sizes. Where the pitch
from one tab ring to the next is roughly constant in concentric labyrinth meshes, and determined only by clearances and
the amount of inter-mesh space necessary for power and data feedthroughs as well as mechanical stability. In offset
meshes, on the other hand, this pitch increases by the offset distance. Even for a small offset this quickly adds up to
an unwieldy total mesh size.

In this section, we conceptually introduce a solution to this problem that allows for larger offsets using a design
where the two meshes interlock like gears. This does mean that the two meshes' rotation must be synchronized, but it
increases the design space of offset labyrinth meshes. For instance, in a gear setup, the wide sides of the inter-mesh
zones can be aligned to lie on the same side, so fiber passthrough can be realized more easily even without the need to
spiral the fiber around the axes of rotation.

\subsection{Mesh synchronization}

For geared meshes to work, both speed and phase of the rotation of the two meshes must be synchronized to a small error.
In this setup, the mesh tabs act like gear teeth. Depending on the ratio between both meshes' tap counts, the two
meshes do not have to rotate at the same rate of rotation and harmonic ratios are possible. Additionally, unlike actual
gears which need to constantly maintain an area of contact, both co-rotating and counter-rotating setups are possible.

\begin{figure}
    \centering
    \subcaptionbox[Offset gear labyrinth mesh assembly render]{\figureattrib{render_side_1.png}}{\includegraphics[width=\textwidth]{\scaledgraphics{render_side_1.png}}}
    \subcaptionbox[Offset gear labyrinth mesh assembly render]{\figureattrib{render_side_2.png}}{\includegraphics[width=\textwidth]{\scaledgraphics{render_side_2.png}}}

    \caption{
        Renderings of the complete offset labyrinth gear mesh assembly.
    }
\end{figure}

\begin{figure}
    \centering
    \includegraphics[width=\textwidth]{gear_plan_1.eps}
    \caption[Offset gear mesh assmbly schema]{\figureattrib{gear_plan_1.svg}}
\end{figure}

\begin{figure}
    \centering
    \includegraphics[width=\textwidth]{gear_plan_2.eps}
    \caption[Offset gear mesh schedule]{\figureattrib{gear_plan_2.svg}}
\end{figure}

\section{Outlook}

\clearpage % clearpage flushes all figures. force this here so we don't get figures floating in between references.
% TODO when breaking this out into a template for building both the whole thesis and individual chapters, we have to
% decide whether we want to keep the bibliography per-chapter or only once for the whole thesis. In the latter case, we
% probably want to replace subbibintoc with bibintoc, or add a custom "bibliography" chapter and adjust the second
% bibliography's heading
\newrefcontext[labelprefix={W}]
\printbibliography[type={online},title={Web sources},heading=subbibintoc]
\newrefcontext
\printbibliography[nottype={online},resetnumbers,heading=subbibintoc]

\appendix

\end{document}