\documentclass{article}

\usepackage{../style}

\usepackage{../langs}

\begin{document}

\fnote{\copyright{} Christian Urban, 2014, 2015}

\section*{Handout 4 (Access Control)}

Access control is essentially about deciding whether to grant

access to a resource or deny it. Sounds easy, no? Well it

turns out that things are not as simple as they seem at first

glance. Let us first look, as a case-study, at how access

control is organised in Unix-like systems (Windows systems

have similar access controls, although the details might be

quite different). Then we have a look at how secrecy and

integrity can be ensured in a system, and finally have a look

at shared access control in multi-agent systems. But before we

start, let us motivate access control systems by the kind of

attacks we have seen in the last lecture. \bigskip

\noindent There are two further general approaches for

countering buffer overflow attacks (and other similar

attacks). One are Unix-like access controls, which enable a

particular architecture for network applications, for example

web-servers. This architecture minimises the attack surface

that is visible from, for example, the Internet. And if an

attack occurs the architecture attempts to limit the damage.

The other approach is to \emph{radically} minimise the attack

surface by running only the bare essentials on the web-server.

In this approach, even the operating system is eliminated.

This approach is called \emph{unikernel}. 

A \emph{unikernel} is essentially a single, fixed purpose

program running on a server. Nothing else is running on the

server, except potentially many instances of this single

program are run concurrently with the help of a

hypervisor.\footnote{Xen is a popular hypervisor; it provides

the mechanism of several virtual machines on a single

computer.} This single program implements the functionality

the server offers (for example serving web-pages). The main

point is that all the services the operating system normally

provides (network stack, file system, ssh and so on) are not

used by default in unikernels. Instead, the single program

uses libraries (the unikernel) whenever some essential

functionality is needed. The developer only needs to select a

minimal set of these libraries in order to implement a server

for web-pages, for example. In this way, ssh, say, is only

provided, when it is absolutely necessary.

Unikernels are a rather recent idea for hardening servers. I

have not seen any production use of this idea, but there are

plenty of examples from academia. The advantage of unikernels

is the rather small footprint in terms of memory, booting

times and so on (no big operating system is needed). This

allows unikernels to run on low-coast hardware such as

Raspberry Pis or Cubieboards, where they can replace much more

expensive hardware for the same purpose. The low booting times

of unikernels are also an advantage when your server needs to

scale up to higher user-demands. Then it is often possible to

just run another instance of the single program, which can be

started almost instantly without the user seeing any delay

(unlike if you have to start, say, Windows and then on top of

that start your network application). One of the most

well-known examples of a unikernel is MirageOS available from

\begin{center}

\url{https://mirage.io}

\end{center}

\noindent This unikernel is based on the functional

programming language Ocaml, which provides added security

(Ocaml does not allow buffer overflow attacks, for example).

If you want to test the security of MirageOS, the 

developers issued a Bitcoin challenge: if you can break into 

their system at

\begin{center}

\url{http://ownme.ipredator.se}

\end{center}

\noindent 

you can get 10 Bitcoins. This is approximately 

However, sometimes you cannot, or do not want to, get rid of

the operating system. In such cases it is still a good idea

to minimise the attack surface. For this it helps if the

network application can be split into two parts---an 

application and an interface:

\begin{center}

  \begin{tikzpicture}[scale=1]

  \draw[line width=1mm] (-.3, 0) rectangle (1.5,2);

  \draw (4.7,1) node {Internet};

  \draw (-2.7,1.5) node {\footnotesize 

  \begin{tabular}{c}Application\\(dangerous part)\end{tabular}};

  \draw (0.6,1.7) node {\footnotesize Interface};

  \draw (0.6,-0.4) node {\footnotesize \begin{tabular}{c}unprivileged\\[-1mm] process\end{tabular}};

  \draw (-2.7,-0.4) node {\footnotesize \begin{tabular}{c}privileged\\[-1mm] process\end{tabular}};

  \draw[line width=1mm] (-1.6, 0) rectangle (-3.8,2);

  \draw[white] (1.7,1) node (X) {};

  \draw[white] (3.7,1) node (Y) {};

  \draw[<->, line width = 2mm] (X) -- (Y);

  \draw[<->, line width = 1mm] (-0.4,1) -- (-1.5,1);

  \end{tikzpicture}

\end{center}

\noindent The idea is that all heavy-duty lifting, or

dangerous operations, in the application (for example database

access or writing a file) is done by a privileged process. All user input from

the internet is received by an \emph{un}privileged process,

which is restricted to only receive user input from the

Internet and communicates with the privileged process. This

communication, however, needs to be sanitised, meaning any

unexpected user-input needs to be rejected. The idea behind

this split is that if an attacker can take control of the

\emph{un}privileged process, then he or she cannot do much

damage. However, the split into such privileged and

unprivileged processes requires an operating system that

supports Unix-style access controls, which look at next.

\subsubsection*{Unix-Style Access Control}

Following the Unix-philosophy that everything is considered as

a file, even memory, ports and so on, access control in Unix

is organised around 11 Bits that specify how a file can be

accessed. These Bits are sometimes called the \emph{permission

attributes} of a file. There are typically three modes for

access: \underline{\textbf{r}}ead, \underline{\textbf{w}}rite

and e\underline{\textbf{x}}ecute. Moreover there are three

user groups to which the modes apply: the owner of the file,

the group the file is associated with and everybody else. 

A typical permission of a file owned by \texttt{bob} 

being in the group \texttt{staff} might look as 

follows:

\begin{center}

${\underbrace{\LARGE\texttt{-}}_{\text{\makebox[0mm]{directory}}}}

 \;{\underbrace{\LARGE\texttt{r{}-{}-}}_{\text{user}}}\,

 {\underbrace{\LARGE\texttt{r{}w{}-}}_{\text{group}}}\,

 {\underbrace{\LARGE\texttt{r{}w{}x}}_{\text{other}}}\;\;\;

 \LARGE\texttt{bob}\;\;\;\texttt{staff}$

\end{center}

\noindent For the moment let us ignore the directory bit. The

Unix access rules imply that Bob will only have read access to

this file, even if he is in the group \texttt{staff} and this

group's access permissions allow read and write. Similarly every

member in the \texttt{staff} group who is not \texttt{bob},

will only have read-write access permissions, not

read-write-execute.

This relatively fine granularity of owner, group, everybody

else seems to cover many useful scenarios of access control. A

typical example of some files with permission attributes is as

follows:

{\small\lstinputlisting[language={}]{../slides/lst}}

\noindent The leading \pcode{d} in Lines 2 and 6 indicate that

the file is a directory, whereby in the Unix-tradition the

\pcode{.} points to the directory itself. The \pcode{..}

points at the directory ``above'', or parent directory. The

second to fourth letter specify how the owner of the file can

access the file. For example Line 3 states that \pcode{ping}

can read and write \pcode{manual.txt}, but cannot execute it.

The next three letters specify how the group members of the

file can access the file. In Line 4, for example, all students

can read and write the file \pcode{report.txt}. Finally the

last three letters specify how everybody else can access a

file. This should all be relatively familiar and

straightforward. No?

There are already some special rules for directories and

links. If the execute attribute of a directory is \emph{not}

set, then one cannot change into the directory and one cannot

access any file inside it. If the write attribute is

\emph{not} set, then one can change existing files (provide

they are changeable), but one cannot create new files. If the

read attribute is \emph{not} set, one cannot search inside the

directory (\pcode{ls -la} does not work) but one can access an

existing file, provided one knows its name. Links to files

never depend on the permission of the link, but the file they

are pointing to. Otherwise one could easily change access

rights to files.

While the above might sound already moderately complicated,

the real complications with Unix-style file permissions

involve the setuid and setgid attributes. For example the file

\pcode{microedit} in Line 5 has the setuid attribute set

(indicated by the \pcode{s} in place of the usual \pcode{x}).

The purpose of setuid and setgid is to solve the following

puzzle: The program \pcode{passwd} allows users to change

their passwords. Therefore \pcode{passwd} needs to have write

access to the file \pcode{/etc/passwd}. But this file cannot

be writable for every user, otherwise anyone can set anyone

else's password. So changing securely passwords cannot be

achieved with the simple Unix access rights discussed so far.

While this situation might look like an anomaly, it is in fact

an often occurring problem. For example looking at current

active processes with \pcode{/bin/ps} requires access to

internal data structures of the operating system, which only

root should be allowed to. In fact any of the following

actions cannot be configured for single users, but need

privileged root access

\begin{itemize}

\item changing system databases (users, groups, routing tables

and so on)

\item opening a network port below 1024

\item interacting with peripheral hardware, such as printers, 

harddisk etc

\item overwriting operating system facilities, like

process scheduling and memory management

\end{itemize}

\noindent This will typically involve quite a lot of programs

on a Unix system. I counted 90 programs with the setuid

attribute set on my bog-standard Mac OSX system (including the

program \pcode{/usr/bin/login}). The problem is that if there

is a security problem with only one of them, be it a buffer

overflow for example, then malicious users can gain root

access (and for outside attackers it is much easier to take

over a system). Unfortunately it is rather easy to cause a

security problem since the handling of elevating and dropping

access rights in such programs rests entirely with the

programmer.

The fundamental idea behind the setuid attribute is that a

file will be able to run not with the callers access rights,

but with the rights of the owner of the file. So

\pcode{/usr/bin/login} will always be running with root access

rights, no matter who invokes this program. The problem is

that this entails a rather complicated semantics of what the

identity of a process (that runs the program) is. One would

hope there is only one such ID, but in fact Unix distinguishes

three(!):

\begin{itemize}

\item \emph{real identity}\\ 

This is the ID of the user who creates 

the process; can only be changed to something else by root. 

\item \emph{effective identity}\\ 

This is the ID that is used to 

grant or deny access to a resource; can be changed to either

the real identity or saved identity by users, can be changed 

to anything by root.

\item \emph{saved identity}\\

If the setuid bit set in a file then the process is started

with the real identity of the user who started the program,

and the identity of the owner of the program as effective and

saved identity. If the setuid bit is not set, then the

saved identity will be the real identity. 

\end{itemize}

\noindent As an example consider again the \pcode{passwd}

program. When started by, say the user \pcode{foo}, it has at

the beginning the identities:

\begin{itemize}

\item \emph{real identity}: \pcode{foo}\\

\emph{effective identity}: \pcode{foo}\\ 

\emph{saved identity}: \pcode{root}

\end{itemize}

\noindent It is then allowed to change the effective

identity to the saved identity to have

\begin{itemize}

\item \emph{real identity}: \pcode{foo}\\

\emph{effective identity}: \pcode{root}\\ 

\emph{saved identity}: \pcode{root}

\end{itemize}

\noindent It can now read and write the file

\pcode{/etc/passwd}. After finishing the job it is supposed to

drop the effective identity back to \pcode{foo}. This is the

responsibility of the programmers who wrote \pcode{passwd}.

Notice that the effective identity is not automatically

elevated to \pcode{root}, but the program itself must make

this change. After it has done the work, the effective

identity should go back to the real identity.

If you want to play more with access rights in Unix, you can

use the program in Figure~\ref{test}. It explicitly checks for

readability and writability of files. The \pcode{main}

function is organised into two parts: the first checks

readability and writability with the permissions according to

a potential setuid bit, and the second (starting in Line 34)

when the permissions are lowered to the caller. Note that this

program has one problem as well: it only gives a reliable

answer in cases a file is {\bf not} readable or {\bf not}

writable when it returns an error code 13 (permission denied).

It sometimes claims a file is not writable, say, but with an

error code 26 (text file busy). This is unrelated to the

permissions of the file.

\begin{figure}[p]

\small

\lstinputlisting[language=C]{../progs/read.c}

\caption{A read/write test program in C. It returns errno = 13 

in cases when permission is denied.\label{test}}

\end{figure}

Despite this complicated semantics, Unix-style access control

is of no use in a number of situations. For example it cannot

be used to exclude some subset of people, but otherwise have

files readable by everybody else (say you want to restrict

access to a file such that your office mates cannot access a

file). You could try setting the group of the file to this

subset and then restrict access accordingly. But this does not

help, because users can drop membership in groups. If one

needs such fine-grained control over who can access a file,

one needs more powerful \emph{mandatory access controls} as

described next.

\subsubsection*{Secrecy and Integrity}

Often you need to keep information secret within a system or

organisation, or secret from the ``outside world''. An example

would be to keep insiders from leaking information to

competitors. The secrecy levels used in the military are an

instance of such an access control system. There you

distinguish usually four secrecy levels:

\begin{itemize}

\item top secret

\item secret

\item confidential

\item unclassified

\end{itemize}

The idea is that the secrets classified as top-secret are most

closely guarded and only accessible to people who have a

special clearance. The unclassified category is the lowest

level not needing any clearance. While the idea behind these

security levels is quite straightforward, there are some

interesting phenomenons that you need to think about when

realising such a system. First this kind of access control

needs to be \emph{mandatory} as opposed to

\emph{discretionary}. With discretionary access control, the

users can decide how to restrict or grant access to resources.

With mandatory access control, the access to resources is

enforced ``system-wide'' and cannot be controlled by the user.

There would be no point to let users set the secrecy level,

because if they want to leak information they would set it to

the lowest. Even if there is no malicious intent, it could

happen that somebody by accident sets the secrecy level too

low for a document. Note also that the secrecy levels are in

tension with the Unix-style access controls. There root is

allowed to do everything, but in a system enforcing secrecy,

you might not like to give root such powers. 

There are also some interesting rules for reading and writing

a resource that need to be enforced: 

\begin{itemize}

\item {\bf Read Rule}: a principal $P$ can read a resource $O$

      provided $P$'s security level is at least as high as

      $O$'s

\item {\bf Write Rule}: a principal $P$ can write a resource

      $O$ provided $O$'s security level is at least as high as

      $P$'s 

\end{itemize} 

\noindent The first rule implies that a principal with secret

clearance can read secret documents or lower, but not

documents classified top-secret. The second rule for writing

needs to be the other way around: someone with secret

clearance can write secret or top-secret documents---no

information is leaked in these cases. In contrast the

principal cannot write confidential documents, because then

information can be leaked to lower levels. These rules about

enforcing secrecy with multi-level clearances are often called

\emph{Bell/LaPadula} model, named after two people who studied

such systems.

A problem with this kind of access control system is when two

people want to talk to each other but are assigned different

security clearances, say secret and confidential. In these

situations, the people with the higher clearance have to lower

their security level and are not allowed to take any document

from the higher level with them to the lower level (otherwise

information could be leaked). In actual systems, this

might mean that people need to log out and log into the system

again---this time with credentials for the lower level.

While secrecy is one property you often want to enforce,

integrity is another. This property ensures that nobody

without adequate clearance can change, or tamper with,

systems. An example for this property is a \emph{fire-wall},

which isolates a local system from threads from the 

Internet, for example. The rule for such a system is

that somebody from inside the fire-wall can write resources

outside the firewall, but you cannot write a resource inside 

the fire-wall from outside. Otherwise an outside can just

tamper with a system in order to break in. In contrast

we can read resources from inside the fire-wall, for example

web-pages. But we cannot read anything from outside the 

fire-wall. Lest we might introduce a virus into the system

(behind the fire-wall). In effect in order to ensure

integrity the read and write rules are reversed from the

case of secrecy:

\begin{itemize}

\item {\bf Read Rule}: a principal $P$ can read a resource $O$

      provided $P$'s security level is lower or equal than

      $O$'s

\item {\bf Write Rule}: a principal $P$ can write a resource

      $O$ provided $O$'s security level is lower or equal than

      $P$'s 

\end{itemize} 

\noindent This kind of access control system is called

\emph{Biba} model, named after Kenneth Biba. Its purpose is to

prevent data modification by unauthorised principals.

The somewhat paradoxical result of the different reading and

writing rules in the \emph{\mbox{Bell}/LaPadula} and

\emph{Biba} models is that we cannot have secrecy and

integrity at the same time in a system, or they need to be

enforced by different means.

\subsubsection*{Multi-Agent Access Control}

In military or banking, for example, very critical decisions

need to be made using a \emph{two-people rule}. This means such

decisions need to be taken by two people together, so that no

single person can defraud a bank or start a nuclear war (you

will know what I mean if you have seen the classic movie ``Dr

Strangelove or: How I Learned to Stop Worrying and Love the

Bomb''\footnote{\url{http://en.wikipedia.org/wiki/Dr._Strangelove}}).

Translating the two-people rule into a software system seems not

as straightforward as one might think.

Let us assume we want to implement a system where CEOs can

make decisions on their own, for example whether or not to

sell assets, but two managing directors (MDs) need to come

together to make the same decision. If ``lowly'' directors

(Ds) want to take this decision, three need to come together.

Remember cryptographic keys are just sequences of bits. A

naive solution to the problem above is to split the necessary

key into $n$ parts according to the ``level'' where the

decision is taken. For example one complete key for a CEO, 

halves of the key for the MDs and thirds for the Ds. The

problem with this kind of sharing a key is that there might be

many hundreds MDs and Ds in your organisations. Simple-minded

halving or devision by three of the key just does not work.

A much more clever solution was proposed by Blakley and Shamir

in 1979. This solution is inspired by some simple geometric

laws. Suppose a three-dimentional axis system. We can, clearly,

specify a point on the $z$-axis, say, by specifying its

coordinates. But we could equally specify this point by a line

that intersects the $z$-axis in this point. How can a line be

specified? Well, by giving two points in space. But as you

might remember from school days, we can specify the point also

by a plane intersecting the $z$-axis and a plane can be

specified by three points in space. This could be pictured as

follows:

\begin{center}

\includegraphics[scale=0.45]{../pics/pointsplane.jpg}

\end{center}

%\begin{center}

%\begin{tikzpicture}
%

%  \draw[->,line width=0.5mm] (0,0,0) -- (2,0,0);
%  \draw[->,line width=0.5mm] (0,0,0) -- (0,2,0);
%  \draw[->,line width=0.5mm] (0,0,0) -- (0,0,2);

%  

%  \path[draw] (-1,-4) to[out=20,in=220] (3,3);

%  \path[draw] (6,-7) to[out=40,in=210] (9,1);

%  \path[draw] (-1,-4) to[out=0,in=80] (6,-7);

%  \path[draw] (3,3) to[out=10,in=140] (9,1);

%
%\end{tikzpicture}

%\end{center}

\noindent The idea is to use the points as keys for each level

of shared access. The CEO gets the point directly. The MDs get

keys lying on a line and the Ds get keys lying on the plane.

Clever, no? Scaling this idea to more dimensions allows for

even more levels of access control and more interesting access

rules, like one MD and 2 Ds can take a decision together.