commit c8a55881b2f7debfa57fd4b17df101f8ceca8aae
Author: Karsten Loesing <karsten.loesing(a)gmx.net>
Date: Sun Oct 14 08:39:23 2012 -0400
Add bridge-stats report from April 2012.
---
2012/bridge-report-usage-stats/.gitignore | 3 +
.../bridge-report-usage-stats.bib | 10 +
.../bridge-report-usage-stats.tex | 217 ++++++++++++++++++++
2012/bridge-report-usage-stats/bridgeusers.png | Bin 0 -> 69368 bytes
2012/bridge-report-usage-stats/discarded.png | Bin 0 -> 78228 bytes
2012/bridge-report-usage-stats/notreported.png | Bin 0 -> 102217 bytes
2012/bridge-report-usage-stats/reported.png | Bin 0 -> 85090 bytes
2012/bridge-report-usage-stats/tortechrep.cls | 1 +
8 files changed, 231 insertions(+), 0 deletions(-)
diff --git a/2012/bridge-report-usage-stats/.gitignore b/2012/bridge-report-usage-stats/.gitignore
new file mode 100644
index 0000000..c787837
--- /dev/null
+++ b/2012/bridge-report-usage-stats/.gitignore
@@ -0,0 +1,3 @@
+bridge-report-usage-stats.pdf
+bridge-report-usage-stats-2012-04-30.pdf
+
diff --git a/2012/bridge-report-usage-stats/bridge-report-usage-stats.bib b/2012/bridge-report-usage-stats/bridge-report-usage-stats.bib
new file mode 100644
index 0000000..289b32b
--- /dev/null
+++ b/2012/bridge-report-usage-stats/bridge-report-usage-stats.bib
@@ -0,0 +1,10 @@
+@techreport{tor-2010-11-001,
+ author = {Sebastian Hahn and Karsten Loesing},
+ title = {Privacy-preserving Ways to Estimate the Number of {Tor} Users},
+ institution = {The Tor Project},
+ number = {2010-11-001},
+ year = {2010},
+ month = {November},
+ url = {https://research.torproject.org/techreports/countingusers-2010-11-30.pdf}
+}
+
diff --git a/2012/bridge-report-usage-stats/bridge-report-usage-stats.tex b/2012/bridge-report-usage-stats/bridge-report-usage-stats.tex
new file mode 100644
index 0000000..969d63a
--- /dev/null
+++ b/2012/bridge-report-usage-stats/bridge-report-usage-stats.tex
@@ -0,0 +1,217 @@
+\documentclass{tortechrep}
+\usepackage{url}
+\usepackage{graphicx}
+\begin{document}
+
+\title{What fraction of our bridges are\\not reporting usage statistics?}
+\author{Karsten Loesing}
+\contact{karsten(a)torproject.org}
+\reportid{2012-04-001}
+\date{April 30, 2012}
+\maketitle
+
+\section{Introduction}
+
+Tor's current approach to count daily bridge users is probably broken.
+The estimate of daily bridge users from all countries ranges between a few
+hundred to half a million in the time between mid-2008 and early 2012 (see
+Figure~\ref{fig:bridgeusers}).
+We have little idea whether the real number is closer to the lower or the
+upper end.
+It's probably ``somewhere in the middle.''
+
+\begin{figure}[t]
+\includegraphics[width=\textwidth]{bridgeusers.png}
+\caption{Estimated bridge users from all countries between 2008 and 2012.}
+\label{fig:bridgeusers}
+\end{figure}
+
+The current approach to estimate the number of bridge users is based on
+bridges reporting the number of unique IP addresses they see in a given
+24-hour timeframe to the bridge authority.
+We collect all reports, sum up unique IP addresses per day, and interpret
+the result as estimated user number.
+
+We already identied two shortcomings in this
+approach~\cite{tor-2010-11-001}: The first shortcoming is that the
+assumption that a bridge user only connects to a single bridge is very
+likely false.
+As a result we may over-count bridge clients connecting to two or more
+bridges.
+The second shortcoming is that we're excluding a yet unknown fraction of
+bridges which don't report usage statistics to the bridge authority.
+A possible reason for not reporting statistics is the minimum uptime of
+24~hours before publishing statistics which has the purpose of hiding
+exact connection times to protect the users' privacy.
+
+In this report we want to focus on the second shortcoming by analyzing
+what fraction of bridges are not reporting usage statistics.
+Obviously, whether this fraction is at 20\% or at 80\% has a major impact
+on the estimated number of bridge users.
+But in addition to that we hope to learn something more general about how
+bridges report statistics to the bridge authority that we can apply to new
+approaches that estimate daily bridge users.
+
+In the following we discuss reasons for discarding reported bridge
+statistics and possible causes for bridges not to report statistics.
+We then look into the bridge descriptor archives to quantify what fraction
+of bridges are affected by these cases.
+We conclude with ideas for increasing the fraction of included statistics.
+
+\section{Reasons for missing bridge usage statistics}
+
+There are two categories of reasons for missing bridge usage statistics:
+either a bridge reports statistics which are discarded, or the bridge does
+not report statistics at all.
+Reasons for discarding reported statistics are:
+
+\begin{enumerate}
+\item \textbf{Running as non-bridge relay:} We exclude all statistics from
+bridges that have been running as non-bridge relays before.
+The reason is that non-bridge clients may still connect to such a bridge.
+We expect there to be many more directly connecting users than bridge
+users, so including these statistics might lead to greatly overestimating
+the number of bridge users.
+We currently exclude statistics from bridges which have been running as
+relay at any time in the past, even months ago.
+We had cases where excluding such a bridge removed a sudden increase in
+bridge user numbers which could not be explained otherwise.
+\item \textbf{Known bug in statistics code:} There are a few Tor versions
+which had bugs in their statistics implementation.
+We exclude these statistics, too.
+\item \textbf{Missing geoip file:} We recently discovered that bridges
+which don't have a geoip file still report bridge usage statistics with
+all zeros.
+For the current approach where we sum up all observations, this isn't a
+problem.
+But it's still interesting to learn how wide-spread the problem of missing
+geoip files on bridges is.
+Only bridges running Tor version 0.2.3.1-alpha or higher report whether
+they have a geoip file congured or not.
+\end{enumerate}
+
+In addition to these cases, there are a few possible causes for bridges
+not reporting statistics:
+
+\begin{enumerate}
+\setcounter{enumi}{3}
+\item \textbf{Less than 24 hours uptime:} Bridges which have an uptime of
+less than 24 hours don't report statistics for this period of time.
+This has to do with the requirement to aggregate observations for a
+sufficient amount of time to hide exact connection times and protect the
+users' privacy.
+\item \textbf{Descriptor publication delay:} Some bridges may even
+complete a 24-hour interval and prepare statistics to be reported in their
+next descriptor.
+But then they go offline and don't publish that descriptor.
+Bridges look at previously finished statistics intervals when starting up,
+but either a bridge decides that its previous statistics are too old to be
+published, or a bridge never shows up again.
+The fix here might be to make bridges publish a new descriptor immediately
+after finishing a statistics interval, which is suggested as enhancement
+\#4142.
+We should probably find out how many bridges are affected by this problem
+before implementing the fix.
+\item \textbf{Other reasons:} There may be other causes for a bridge not
+reporting statistics which we did not identify.
+\end{enumerate}
+
+\section{Fraction of missing bridge usage statistics}
+
+After listing reasons for reported observations being discarded and for
+bridges not reporting statistics at all, we now want to quantify how many
+bridges are affected by which case.
+
+\begin{figure}[t]
+\includegraphics[width=\textwidth]{reported.png}
+\caption{Fraction of bridges that reported statistics which were either
+used or discarded, or that did not report statistics.}
+\label{fig:reported}
+\end{figure}
+
+Figure~\ref{fig:reported} shows the fraction of bridges that did or did
+not report usage statistics and how many of these reports had to be
+discarded.
+The graph shows an almost monotonic downward trend of non-reported
+statistics from 2008 to early 2010 to around 20\%.
+This fraction went up only slightly in 2010 and 2011 to 40\% and is now
+back at 20\%.
+The fraction of reported and discarded statistics was between 10\% and
+20\% for most of the time between 2008 and today.
+As a result, the fraction of reported and used statistics went up from
+around 35\% in early 2009 to around 75\% in early 2012.
+
+These results are much better than expected before.
+A bridge usage statistic that is based on 75\% of all bridges at least
+rules out inaccuracies from too little sample sizes.
+We still want to look into reasons for discarded or not reported
+statistics.
+
+\begin{figure}[t]
+\includegraphics[width=\textwidth]{discarded.png}
+\caption{Reasons for discarding reported usage statistics.}
+\label{fig:discarded}
+\end{figure}
+
+Figure~\ref{fig:discarded} shows what fractions of reported bridge
+statistics were discarded for what reasons.
+The fractions of statistics that had to be discarded because of the
+geoip-stats bug in Tor 0.2.2.x or because of missing geoip files are at
+almost 0\% for most of the time.
+Only in late 2009, the geoip-stats bug affected up to 5\% of bridges.
+But the fraction of discarded statistics because of bridges previously
+running as non-bridge relays is quite high at 10\% to 20\%.
+
+It's quite likely that we could reduce this fraction by being less strict
+about bridges running as non-bridge relays.
+In theory, a delay of a few days between running as relay and running as
+bridge should be sucient to exclude directly connecting clients from the
+statistics.
+However, this requires further analysis.
+
+\begin{figure}[t]
+\includegraphics[width=\textwidth]{notreported.png}
+\caption{Reasons for bridges not reporting usage statistics.}
+\label{fig:notreported}
+\end{figure}
+
+Figure~\ref{fig:notreported} shows what fractions of bridges don't report
+statistics for what reasons.
+For most of the time, the fraction of bridges not reporting statistics
+because they went online before their 24-hour interval ended was about as
+large as the fraction of ``other reasons.''
+Only recently, the ``other reasons'' dropped to almost 0\%.
+The fraction of missing statistics due to a delay between completing a
+statistics interval and publishing the descriptor containing those
+statistics is almost at 0\% for most of the time.
+There's probably not much that we can do about the first category of
+bridges which go online before their 24-hour interval ends.
+This interval is there to hide exact connection times and to protect the
+users' privacy.
+Any algorithm will have to cope with 15\% to 25\% missing statistics due
+to the 24-hour interval requirement.
+Fortunately, the fraction of non-reported statistics due to the descriptor
+publication delay is almost at 0\%, so we don't have to fix that.
+It's unclear what other reasons led to bridges not publishing statistics.
+Given that this fraction is almost at 0\%, there's no immediate need to
+investigate.
+
+\section{Conclusion}
+
+In this report we analyzed what fraction of bridges are not reporting
+usage statistics, which might affect our daily bridge user estimates.
+The analysis of bridge descriptor archives resulted in a fraction of up to
+75\% of bridges reporting usage statistics that get used to estimate user
+numbers.
+This fraction might even be increased by discarding fewer statistics from
+bridges that were seen as non-bridge relays before.
+We conclude that a too small sample size is not the issue of our probably
+wrong bridge user numbers.
+We think that a new approach that will be based on bridges reporting their
+findings in 24-hour intervals has a good chance of leading to quite
+reliable user numbers.
+
+\bibliography{bridge-report-usage-stats}
+
+\end{document}
+
diff --git a/2012/bridge-report-usage-stats/bridgeusers.png b/2012/bridge-report-usage-stats/bridgeusers.png
new file mode 100755
index 0000000..a77e87a
Binary files /dev/null and b/2012/bridge-report-usage-stats/bridgeusers.png differ
diff --git a/2012/bridge-report-usage-stats/discarded.png b/2012/bridge-report-usage-stats/discarded.png
new file mode 100755
index 0000000..a96a604
Binary files /dev/null and b/2012/bridge-report-usage-stats/discarded.png differ
diff --git a/2012/bridge-report-usage-stats/notreported.png b/2012/bridge-report-usage-stats/notreported.png
new file mode 100755
index 0000000..d44f352
Binary files /dev/null and b/2012/bridge-report-usage-stats/notreported.png differ
diff --git a/2012/bridge-report-usage-stats/reported.png b/2012/bridge-report-usage-stats/reported.png
new file mode 100755
index 0000000..fa5640f
Binary files /dev/null and b/2012/bridge-report-usage-stats/reported.png differ
diff --git a/2012/bridge-report-usage-stats/tortechrep.cls b/2012/bridge-report-usage-stats/tortechrep.cls
new file mode 120000
index 0000000..4c24db2
--- /dev/null
+++ b/2012/bridge-report-usage-stats/tortechrep.cls
@@ -0,0 +1 @@
+../../tortechrep.cls
\ No newline at end of file
