learning_setup.tex 20.2 KB
Newer Older
Paul Fiterau Brostean's avatar
updates    
Paul Fiterau Brostean committed
1
\section{The learning setup} \label{sec:setup}
2
3
4
5
6
7
%This chapter will cover the setup used to infer the state machines. We provide a general setup outline in Section~\ref{components}. The tested SSH servers are described in Section~\ref{suts}, which were queried with the alphabet described in Section~\ref{alphabet}. Section~\ref{setup-handling} will cover the challenging SUT behaviour faced when implementing the mapper, and the adaptations that were made to overcome these challenges. Section~\ref{layers-individual} will discuss the relation between state machines for individual layers and the state machine of the complete SSH protocol. The conventions on visualisation of the inferred state machines are described in Section~\ref{visualisation}.

%Throughout this chapter, an individual SSH message to a SUT is denoted as a \textit{query}. A \textit{trace} is a sequence of multiple queries, starting from a SUT's initial state. Message names in this chapter are usually self-explanatory, but a mapping to the official RFC names is provided in Appendix~\ref{appendixa}.

%\section{Components}\label{components}

8
The learning setup consists of three components: the {\dlearner}, the {\dmapper} and the {\dsut}. The {\dlearner} generates abstract inputs, representing SSH messages. The {\dmapper} transforms these messages into well-formed SSH packets and sends them to the {\dsut}. The {\dsut} sends response packets back to the {\dmapper}, which in turn, translates these packets to abstract outputs. The {\dmapper} then sends the abstract outputs back to the {\dlearner}. 
9
10


Erik Poll's avatar
shit    
Erik Poll committed
11
The {\dlearner} uses LearnLib ~\cite{LearnLib2009}, a Java library implementing $L^{\ast}$ based algorithms for learning Mealy machines. The {\dmapper} is based on Paramiko, an open source SSH implementation written in Python\footnote{Paramiko is available at \url{http://www.paramiko.org/}}. We opted for Paramiko because its code is relatively well structured and documented. The {\dsut} can be any existing implementation of an SSH server. The three components communicate over sockets, as shown in Figure~\ref{fig:components}.
12
\begin{figure}
13
	\centering
14
  \includegraphics[scale=0.29]{components_cropped.pdf}
Paul Fiterau Brostean's avatar
updates    
Paul Fiterau Brostean committed
15
  \caption{The SSH learning setup.}
16
17
18
  \label{fig:components}
\end{figure}

Erik Poll's avatar
shit    
Erik Poll committed
19
20
21
22
SSH is a complex client-server protocol. In our work so far we concentrated on learning models of the implementation of the server, and not of the client.
We further restrict learning to only exploring the terminal service of the Connection layer, as we consider it to be the most interesting
from a security perspective. Algorithms for encryption, compression and hashing are left to default settings and are not purposefully explored. Also, the starting
state of the {\dsut} is one where a TCP connection has already been established and where SSH versions have been exchanged, which are prerequisites for starting the Transport layer protocol.
23

24
25
%figure
%It is therefore important to focus on messages for which interesting state-changing behaviour can be expected. 
Erik Poll's avatar
shit    
Erik Poll committed
26

27
\subsection{The learning alphabet}\label{subsec:alphabet}
28

Paul Fiterau Brostean's avatar
Paul Fiterau Brostean committed
29

Erik Poll's avatar
shit    
Erik Poll committed
30
31
The alphabet we use consists of inputs, which correspond to messages
sent to the server, and outputs, which correspond to messages received
Paul Fiterau Brostean's avatar
Paul Fiterau Brostean committed
32
from the server. We split \emph{the input alphabet} into three parts, one
33
for each of the protocol layers. 
Paul Fiterau Brostean's avatar
Paul Fiterau Brostean committed
34

35
36
%\marginpar{\tiny Erik: the output alphabet is not discussed anywhere,
%but for the discussion of the mapper in the next section it should be}
Paul Fiterau Brostean's avatar
Paul Fiterau Brostean committed
37
Learning does not scale with a growing input alphabet, and since we are only
Erik Poll's avatar
shit    
Erik Poll committed
38
39
40
41
42
learning models of servers, we remove those inputs that are not
intended to ever be sent to the server\footnote{This means we exclude
the messages \textsc{service\_accept}, \textsc{ua\_accept},
\textsc{ua\_failure}, \textsc{ua\_banner}, \textsc{ua\_pk\_ok},
\textsc{ua\_pw\_changereq}, \textsc{ch\_success} and
Erik Poll's avatar
Erik Poll committed
43
\textsc{ch\_failure} from our alphabet.}. Furthermore, from the
Erik Poll's avatar
shit    
Erik Poll committed
44
45
Connection layer we only use messages for channel management and the
terminal functionality.  Finally, because we will only explore
46
protocol behavior after SSH versions have been exchanged, we exclude
Erik Poll's avatar
Erik Poll committed
47
the messages for exchanging version numbers.  
Paul Fiterau Brostean's avatar
Paul Fiterau Brostean committed
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63

%The alphabet comprises abstract messages corresponding to the concrete messages exchanged during client/server interactions, as per RFC. All included messages follow
%the Byte Protocol format, and all are processable, that is, the SSH server should react to all of them. Hence, all messages could be
%included as inputs in the alphabet. Doing so, however, would lead to prohibitively 
%long learning times. Consequently, we restrict the input alphabet to messages a server
%is expected to receive from the client. \footnote{Which means we exclude
%the messages \textsc{service\_accept}, \textsc{ua\_accept},
%\textsc{ua\_failure}, \textsc{ua\_banner}, \textsc{ua\_pk\_ok},
%\textsc{ua\_pw\_changereq}, \textsc{ch\_success} and
%\textsc{ch\_failure} from our input alphabet.} 
%
%
%For the same reason, from the Connection layer we only use as inputs messages for channel management and the
%terminal functionality.  Finally, as we only explore protocol behavior after SSH versions have been exchanged, we exclude
%the messages for exchanging version numbers.  We don't place any restrictions on the 
%output of the server. All outputs the server generates are also included in the output alphabet. 
64

Erik Poll's avatar
shit    
Erik Poll committed
65
The resulting lists of inputs for the three protocol layers are given
Paul Fiterau Brostean's avatar
Paul Fiterau Brostean committed
66
in Tables~\ref{trans-alphabet}-\ref{conn-alphabet}.  In some
Erik Poll's avatar
shit    
Erik Poll committed
67
68
69
70
71
72
experiments, we used only a subset of the most essential inputs, to
further speed up experiments. This \textit{restricted alphabet}
significantly decreases the number of queries needed for learning
models while only marginally limiting explored behavior.  We discuss
this again in Section~\ref{sec:result}. Inputs included in the
restricted alphabet are marked by '*' in the tables below.
73

Erik Poll's avatar
Erik Poll committed
74
Table~\ref{trans-alphabet} lists the Transport layer inputs. We include a version of the \textsc{kexinit} message with \texttt{first\_kex\_packet\_follows} disabled.
75
This means no guess~\cite[p. 17]{rfc4253} is attempted on the {\dsut}'s parameter preferences. Consequently, the {\dsut} will have to send its own \textsc{kexinit} in order to 
Paul Fiterau Brostean's avatar
Paul Fiterau Brostean committed
76
convey its own parameter preferences before key exchange can proceed. Also included are inputs for establishing new keys (\textsc{kex30}, \textsc{newkeys}), disconnecting (\textsc{disconnect}), as well as the special inputs \textsc{ignore}, \textsc{unimpl} and \textsc{debug}. The latter are not interesting, as they are normally ignored by implementations. Hence they are excluded from our restricted alphabet. \textsc{disconnect} proved costly time wise, so was also excluded.
77
78
%We include two versions of the \textsc{kexinit} message, one where \texttt{first\_kex\_packet\_follows} is disabled, the other when it is enabled, in which case, the message would make a guess on the security parameters~\cite[p. 17]{rfc4253}. Our mapper can only handle correct key guesses, so the wrong-guess procedure as described in ~\cite[p. 19]{rfc4253} was not supported. 
%\textsc{ignore}, \textsc{unimpl} and \textsc{debug} 
79
%When needed, SUTs were configured to make this guess work by altering their cipher preferences. The SSH version and comment string (described in Section~\ref{ssh-run-trans}) was not queried because it does not follow the binary packet protocol.
80
81
82

\begin{table}[!ht]
\centering
83
\small
84
85
\begin{tabular}{ll}
\textbf{Message} & \textbf{Description} \\
86
\textsc{disconnect} & Terminates the current connection~\cite[p. 23]{rfc4253} \\
87
\textsc{ignore} & Has no intended effect~\cite[p. 24]{rfc4253} \\
88
\textsc{unimpl} & Intended response to unrecognized messages~\cite[p. 25]{rfc4253} \\
89
\textsc{debug} & Provides other party with debug information~\cite[p. 25]{rfc4253} \\
90
\textsc{kexinit}* & Sends parameter preferences~\cite[p. 17]{rfc4253} \\
91
%\textsc{guessinit}* & A \textsc{kexinit} after which a guessed \textsc{kex30} follows~\cite[p. 19]{rfc4253} \\
92
93
94
95
\textsc{kex30}* & Initializes the Diffie-Hellman key exchange~\cite[p. 21]{rfc4253} \\
\textsc{newkeys}* & Requests to take new keys into use~\cite[p. 21]{rfc4253} \\
\textsc{sr\_auth}* & Requests the authentication protocol~\cite[p. 23]{rfc4253} \\
\textsc{sr\_conn}* & Requests the connection protocol~\cite[p. 23]{rfc4253}
96
\end{tabular}
97
\caption{Transport layer inputs}
98
\vspace{-7mm}
99
100
101
\label{trans-alphabet}
\end{table}

Paul Fiterau Brostean's avatar
Paul Fiterau Brostean committed
102
The Authentication layer defines one single client message type in the form of the authentication request~\cite[p. 4]{rfc4252}. Its parameters contain all information needed for authentication. Four authentication methods exist: none, password, public key and host-based. Our mapper supports all methods except the host-based authentication because some SUTs don't support this feature. Both the public key and password methods have  \textsc{ok} and \textsc{nok} variants, which provide respectively correct and incorrect credentials. Our restricted alphabet supports only public key authentication, as the implementations processed this faster than the other authentication methods.
103
104
105

\begin{table}[!ht]
\centering
106
\small
107
108
109
\begin{tabular}{ll}
\textbf{Message} & \textbf{Description} \\
\textsc{ua\_none} & Authenticates with the ``none'' method~\cite[p. 7]{rfc4252} \\
110
111
112
113
\textsc{ua\_pk\_ok}* & Provides a valid name/key pair~\cite[p. 8]{rfc4252} \\
\textsc{ua\_pk\_nok}* & Provides an invalid name/key pair~\cite[p. 8]{rfc4252} \\
\textsc{ua\_pw\_ok} & Provides a valid name/password pair~\cite[p. 10]{rfc4252} \\
\textsc{ua\_pw\_nok} & Provides an invalid name/password pair~\cite[p. 10]{rfc4252} \\
114
\end{tabular}
115
\caption{Authentication layer inputs}
116
\vspace{-7mm}
117
118
119
\label{auth-alphabet}
\end{table}

120
The Connection layer allows the client to manage channels and to request/run services over them. In accordance with our learning goal,
121
our mapper only supports inputs for requesting terminal emulation, plus inputs for channel management as shown in Table~\ref{conn-alphabet}.
122
The restricted alphabet only supports the most general channel management inputs. Those excluded are not expected to produce state change.
123

124
125
126

\begin{table}[!ht]
\centering
127
\small
128
129
\begin{tabular}{ll}
\textbf{Message} & \textbf{Description} \\
Paul Fiterau Brostean's avatar
Paul Fiterau Brostean committed
130
131
132
133
\textsc{ch\_open}* & Opens a new channel~\cite[p. 5]{rfc4254} \\
\textsc{ch\_close}* & Closes a channel~\cite[p. 9]{rfc4254} \\
\textsc{ch\_eof}* & Indicates that no more data will be sent~\cite[p. 9]{rfc4254} \\
\textsc{ch\_data}* & Sends data over the channel~\cite[p. 7]{rfc4254} \\
134
135
\textsc{ch\_edata} & Sends typed data over the channel~\cite[p. 8]{rfc4254} \\
\textsc{ch\_window\_adjust} & Adjusts the window size~\cite[p. 7]{rfc4254} \\
Paul Fiterau Brostean's avatar
Paul Fiterau Brostean committed
136
\textsc{ch\_request\_pty}* & Requests terminal emulation~\cite[p. 11]{rfc4254} \\
137
\end{tabular}
138
\caption{Connection layer inputs}
139
\vspace{-7mm}
140
141
\label{conn-alphabet}
\end{table}
Paul Fiterau Brostean's avatar
Paul Fiterau Brostean committed
142
143
144

\emph{The output alphabet} subsumes all messages an SSH server generates, which may include, with identical meaning, any of the messages defined as inputs. They also include responses to various requests: \textsc{kex31}~\cite[p. ]{} as reply to \textsc{kex30}, \textsc{sr\_succes} in response to service requests (\textsc{sr\_auth} and \textsc{sr\_conn}), \textsc{ua\_success} and \textsc{ua\_failure} in response to authentication requests, and \textsc{ch\_open\_success}, \textsc{ch\_success} and \textsc{ch\_close}, which confirm the requested channel operation. To these outputs, we add \textsc{no\_resp} for when the {\dsut} generates no output, and the special outputs \textsc{ch\_none}, \textsc{ch\_max} and \textsc{no\_conn}, and \textsc{buffered}, which we discuss in the next Subsections.

145
%The learning alphabet comprises of input/output messages by which the {\dlearner} interfaces with the {\dmapper}. Section~\ref{sec:ssh} outlines essential inputs, while Table X provides a summary
Paul Fiterau Brostean's avatar
updates    
Paul Fiterau Brostean committed
146
%of all messages available at each layer. \textit{\textit{}}
147

148
%table
149

150
\subsection{The mapper}\label{subsec:mapper}
Paul Fiterau Brostean's avatar
updates    
Paul Fiterau Brostean committed
151

Erik Poll's avatar
shit    
Erik Poll committed
152
153
The {\dmapper} must provide a translation between abstract messages
and well-formed SSH messages: it has to translate abstract inputs
154
listed in Tables~\ref{trans-alphabet}-\ref{conn-alphabet} to actual
Erik Poll's avatar
shit    
Erik Poll committed
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
SSH packets, and translate the SSH packets received in answer
to our abstract outputs.

A special case here occurs when no output is received from the
{\dsut}; in that case the {\dmapper} gives back to the learner a
\textsc{no\_resp} message, to indicate that a time-out occurred.

The sheer complexity of the {\dmapper} meant that it was easier to
adapt an existing SSH implementation, rather than construct the
{\dmapper} from scratch. Paramiko already provides mechanisms for
encryption/decryption, as well as routines for constructing and
sending the different types of packets, and for receiving them. These
routines are called by control logic dictated by Paramiko's own state
machine.  The {\dmapper} was constructed by replacing this control
logic with one dictated by messages received from the {\dlearner}.
%over a socket connection

The {\dmapper} maintains a set of state variables to record parameters
Paul Fiterau Brostean's avatar
Paul Fiterau Brostean committed
173
of the ongoing session, including the server's preferences
Erik Poll's avatar
shit    
Erik Poll committed
174
for key exchange and encryption algorithm, parameters of these
175
protocols, and -- once it has been established -- the session key.
Erik Poll's avatar
shit    
Erik Poll committed
176
177
These parameters are updated when receiving messages from the server,
and are used to concretize inputs to actual SSH messages to the server.
Paul Fiterau Brostean's avatar
updates    
Paul Fiterau Brostean committed
178

Paul Fiterau Brostean's avatar
Paul Fiterau Brostean committed
179
For example, upon receiving a \textsc{kexinit} from the {\dsut}, the {\dmapper} saves
Erik Poll's avatar
shit    
Erik Poll committed
180
181
the {\dsut}'s preferences for key exchange, hashing and encryption
algorithms. Initially these parameters are all set to the defaults
Paul Fiterau Brostean's avatar
Paul Fiterau Brostean committed
182
183
184
185
186
187
188
that any server should support, as required by the RFC. On receiving \textsc{kex31} 
in response to the \textsc{kex30} input, the {\dmapper} saves the hash, as well as the new
keys. Finally, a \textsc{newkeys} response prompts the {\dmapper} to use the new keys 
negotiated earlier  in place of the older ones, if such existed.


The {\dmapper} also contains a buffer for storing opened channels, which is initially empty.
189
190
On a \textsc{ch\_open} from the learner, the {\dmapper} adds a channel to the buffer
with a randomly generated channel identifier, on a \textsc{ch\_close}, it removes the channel
Paul Fiterau Brostean's avatar
Paul Fiterau Brostean committed
191
(if there was any). The buffer size, or the maximum number of opened channels, is limited to one.  Initially the buffer is empty. The {\dmapper} also stores the sequence number of  the last received message from the {\dsut}. This number is then used when constructing \textsc{unimpl} inputs. 
Erik Poll's avatar
shit    
Erik Poll committed
192
193

In the following cases, inputs are answered by the {\dmapper} directly
194
instead of being sent to the {\dsut} to find out its response: (1) on receiving a \textsc{ch\_open} input if the buffer has reached the size limit, the {\dmapper} directly responds with \textsc{ch\_max}; (2)
Paul Fiterau Brostean's avatar
Paul Fiterau Brostean committed
195
196
197
198
199
200
201
202
203
204
205
206
on receiving any input operating on a channel (all Connection layer inputs other than \textsc{ch\_open}) when the buffer is empty, the
{\dmapper} directly responds with \textsc{ch\_none}; (3) if connection with the {\dsut} was terminated, the {\dmapper}
responds with a \textsc{no\_conn} message, as sending further messages to the {\dsut} is pointless in that case.

%\begin{enumerate}
%\item on receiving a \textsc{ch\_open} input and the buffer has reached the size limit, the {\dmapper} directly responds with \textsc{ch\_max};
%\item on receiving any input operating on a channel (all Connection layer inputs other than \textsc{ch\_open}) when the buffer is empty, the
%{\dmapper} directly responds with \textsc{ch\_none};
%\item if connection with the {\dsut} was terminated, the {\dmapper}
%     responds with a \textsc{no\_conn} message, as sending further
%     messages to the {\dsut} is pointless in that case;
%\end{enumerate}
Erik Poll's avatar
Erik Poll committed
207
208
209
%
In many ways, the {\dmapper} acts similar to an SSH client, hence the
decision to built it by adapting an existing client implementation.
210
211


Erik Poll's avatar
Erik Poll committed
212
\subsection{Practical complications}
Erik Poll's avatar
shit    
Erik Poll committed
213

Erik Poll's avatar
Erik Poll committed
214
215
216
217
218
219
%There are three practical complications in learning models of SSH
%servers: (1) an SSH server may exhibit \emph{non-determistic}
%behaviour; (2) a single input to the server can produce a
%\emph{sequence} of outputs ratheer than just a single output, and (3)
%\emph{buffering} behaviour of the server. These complication are
%discussed below.
Paul Fiterau Brostean's avatar
Paul Fiterau Brostean committed
220
221
222
SSH implementations even behind the {\dmapper} abstraction may not behave
like deterministic Mealy Machines, a prerequisite for the learning algorithm
to succeed. Sources of non-determinism are:
Erik Poll's avatar
shit    
Erik Poll committed
223

224
\begin{enumerate}
Erik Poll's avatar
shit    
Erik Poll committed
225
\item Underspecification in the SSH specification (for example, by not
Erik Poll's avatar
Erik Poll committed
226
     specifying the order of certain messages) allows some
227
     non-deterministic behavior. Even if client
Erik Poll's avatar
shit    
Erik Poll committed
228
     and server do implement a fixed order for messages they sent, the
Erik Poll's avatar
Erik Poll committed
229
230
     asynchronous nature of communication means that the
     interleaving of sent and received messages may vary.  Moreover,
Erik Poll's avatar
shit    
Erik Poll committed
231
     client and server are free to intersperse \textsc{debug} and
Erik Poll's avatar
Erik Poll committed
232
233
     \textsc{ignore} messages at any given time\footnote{The \textsc{ignore}
     messages are aimed to thwart traffic analysis.}
234
\item Timing is another source of non-deterministic behavior. For
Erik Poll's avatar
shit    
Erik Poll committed
235
     example, the {\dmapper} might time-out before the {\dsut} had
236
     sent its response. Some {\dsuts} also behave 
Erik Poll's avatar
shit    
Erik Poll committed
237
     unexpectedly when a new query is received too shortly after the
238
239
240
241
242
243
244
245
246
     previous one. Hence in our experiments we adjusted time-out 
		 periods accordingly so that neither of these events occur, and the {\dsut}
		 behaves deterministically all the time.
     %did not occur.  
		
		%However, other timing-related quirks can still
     %cause non-determinism. For example, some {\dsuts} behave
     %unexpectedly when a new query is received too shortly after the
     %previous one.
247
248
%For example, a trace in which a valid user authentication is performed within five milliseconds after an authentication request on DropBear can cause the authentication to (wrongly) fail.  
\end{enumerate}
Erik Poll's avatar
shit    
Erik Poll committed
249
%
Erik Poll's avatar
Erik Poll committed
250
To detect non-determinism, the {\dmapper} caches all observations
Paul Fiterau Brostean's avatar
Paul Fiterau Brostean committed
251
252
253
254
255
256
257
in an SQLite database and verifies if a new observations are consistent
with previous ones.  If not, it raises a warning, which then needs to be manually investigated. We acted on
each case until we found a setting under which behavior was deterministic.

The cache also acts as a cheap source of responses for already answered queries.
Finally, by re-loading the cache from a previous experiment, we were able to start
from where this experiment left off. This proved useful, as experiments could take several days.
Erik Poll's avatar
Erik Poll committed
258

Paul Fiterau Brostean's avatar
Paul Fiterau Brostean committed
259
260
261
262
263
264
%An added benefit of the cache is that it allows the {\dmapper} to
%supply answer to some inputs without actually sending them to the
%{\dsut}. This sped up learning a lot when we had to restart
%experiments: any new experiment on the same {\dsut} could start where
%the previous experiment left off, without re-running all inputs.  This
%was an important benefit, as experiments could take several days.
Paul Fiterau Brostean's avatar
Paul Fiterau Brostean committed
265
266

%A subsequent identical learning run can quickly resume from where the previous one was ended, as the cache from the previous run is used to quickly respond to all queries up to the point the previous run ended.
267

268
Another practical problem besides non-determinism is that an SSH server
Erik Poll's avatar
Erik Poll committed
269
270
may produce a sequence of outputs in response to a single input. This
means it is not behaving as a Mealy machines, which allows for
Paul Fiterau Brostean's avatar
Paul Fiterau Brostean committed
271
only one output. To deal with this, the {\dmapper}
Erik Poll's avatar
Erik Poll committed
272
273
274
275
276
concatenates all outputs into one, and it produces this sequence as
the single output to the {\dlearner}. 

A final challenge is presented by forms of `buffering', which we
encountered in two situations.  Firstly, some implementations buffer
Paul Fiterau Brostean's avatar
Paul Fiterau Brostean committed
277
incoming requests during rekey; only once rekeying is
Erik Poll's avatar
Erik Poll committed
278
279
280
281
282
complete are all these messages processed. This leads to a
\textsc{newkeys} response (indicating rekeying has completed),
directly followed by all the responses to the buffered requests.  This
would lead to non-termination of the learning algorithm, as for every
sequence of buffered messages the response is different.  To
Paul Fiterau Brostean's avatar
Paul Fiterau Brostean committed
283
prevent this, we treat the sequence of queued responses as the single
Erik Poll's avatar
Erik Poll committed
284
285
output \textsc{buffered}.

Paul Fiterau Brostean's avatar
Paul Fiterau Brostean committed
286
A different form of buffering occurs when opening and closing channels, since a
Erik Poll's avatar
Erik Poll committed
287
{\dsut} can close only as many channels as have previously been opened.
288
Learning this behavior would lead to an infinite state machine, as we
Erik Poll's avatar
Erik Poll committed
289
290
would need a state `there are $n$ channels open' for every number $n$.
For this reason, we restrict the number of simultaneously open
Erik Poll's avatar
shit    
Erik Poll committed
291
292
293
channels to one. The {\dmapper} returns a custom response
\textsc{ch\_max} to a \textsc{ch\_open} message whenever this limit is
reached.