[tor-commits] [tech-reports/master] Add new report on counting daily bridge users.
karsten at torproject.org
karsten at torproject.org
Wed Oct 24 15:11:51 UTC 2012
commit 87981946bb08a21fd247a1ea344fc8e36533e382
Author: Karsten Loesing <karsten.loesing at gmx.net>
Date: Wed Oct 24 11:11:11 2012 -0400
Add new report on counting daily bridge users.
---
2012/counting-daily-bridge-users/.gitignore | 3 +
2012/counting-daily-bridge-users/bridgesets.pdf | Bin 0 -> 26476 bytes
.../compare-totalusers-q3-2012.pdf | Bin 0 -> 6366 bytes
.../compare-totalusers.pdf | Bin 0 -> 10156 bytes
2012/counting-daily-bridge-users/consensuses.pdf | Bin 0 -> 6512 bytes
.../counting-daily-bridge-users.bib | 30 ++
.../counting-daily-bridge-users.tex | 533 ++++++++++++++++++++
.../extrapolated-responses.pdf | Bin 0 -> 13134 bytes
.../responses-single-bridges.png | Bin 0 -> 37038 bytes
2012/counting-daily-bridge-users/responses.pdf | Bin 0 -> 7211 bytes
2012/counting-daily-bridge-users/syusers.pdf | Bin 0 -> 7460 bytes
2012/counting-daily-bridge-users/tortechrep.cls | 1 +
.../totalusers-q3-2012.pdf | Bin 0 -> 5307 bytes
2012/counting-daily-bridge-users/totalusers.pdf | Bin 0 -> 7185 bytes
14 files changed, 567 insertions(+), 0 deletions(-)
diff --git a/2012/counting-daily-bridge-users/.gitignore b/2012/counting-daily-bridge-users/.gitignore
new file mode 100644
index 0000000..b36a48e
--- /dev/null
+++ b/2012/counting-daily-bridge-users/.gitignore
@@ -0,0 +1,3 @@
+counting-daily-bridge-users.pdf
+counting-daily-bridge-users-2012-10-24.pdf
+
diff --git a/2012/counting-daily-bridge-users/bridgesets.pdf b/2012/counting-daily-bridge-users/bridgesets.pdf
new file mode 100644
index 0000000..2328f54
Binary files /dev/null and b/2012/counting-daily-bridge-users/bridgesets.pdf differ
diff --git a/2012/counting-daily-bridge-users/compare-totalusers-q3-2012.pdf b/2012/counting-daily-bridge-users/compare-totalusers-q3-2012.pdf
new file mode 100644
index 0000000..5539e94
Binary files /dev/null and b/2012/counting-daily-bridge-users/compare-totalusers-q3-2012.pdf differ
diff --git a/2012/counting-daily-bridge-users/compare-totalusers.pdf b/2012/counting-daily-bridge-users/compare-totalusers.pdf
new file mode 100644
index 0000000..180a6ea
Binary files /dev/null and b/2012/counting-daily-bridge-users/compare-totalusers.pdf differ
diff --git a/2012/counting-daily-bridge-users/consensuses.pdf b/2012/counting-daily-bridge-users/consensuses.pdf
new file mode 100644
index 0000000..916711d
Binary files /dev/null and b/2012/counting-daily-bridge-users/consensuses.pdf differ
diff --git a/2012/counting-daily-bridge-users/counting-daily-bridge-users.bib b/2012/counting-daily-bridge-users/counting-daily-bridge-users.bib
new file mode 100644
index 0000000..1037cbc
--- /dev/null
+++ b/2012/counting-daily-bridge-users/counting-daily-bridge-users.bib
@@ -0,0 +1,30 @@
+ at techreport{tor-2012-04-001,
+ author = {Karsten Loesing},
+ title = {What fraction of our bridges are not reporting usage statistics?},
+ institution = {The Tor Project},
+ number = {2012-04-001},
+ year = {2012},
+ month = {April},
+ note = {\url{https://research.torproject.org/techreports/bridge-report-usage-stats-2012-04-30.pdf}},
+}
+
+ at 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},
+ note = {\url{https://research.torproject.org/techreports/countingusers-2010-11-30.pdf}},
+}
+
+ at inproceedings{loesing2010case,
+ title = {A Case Study on Measuring Statistical Data in the {Tor} Anonymity Network},
+ author = {Karsten Loesing and Steven J. Murdoch and Roger Dingledine},
+ booktitle = {Proc.\ Workshop on Ethics in Computer Security Research},
+ year = {2010},
+ month = Jan,
+ address = {Tenerife, Canary Islands, Spain},
+ note = {\url{https://metrics.torproject.org/papers/wecsr10.pdf}},
+}
+
diff --git a/2012/counting-daily-bridge-users/counting-daily-bridge-users.tex b/2012/counting-daily-bridge-users/counting-daily-bridge-users.tex
new file mode 100644
index 0000000..81eca55
--- /dev/null
+++ b/2012/counting-daily-bridge-users/counting-daily-bridge-users.tex
@@ -0,0 +1,533 @@
+\documentclass{tortechrep}
+\usepackage{url}
+\usepackage{graphicx}
+\begin{document}
+
+\title{Counting daily bridge users}
+\author{Karsten Loesing}
+\contact{karsten at torproject.org}
+\reportid{2012-10-001}
+\date{October 24, 2012}
+\maketitle
+
+\begin{abstract}
+As part of the Tor Metrics Project, we want to learn how many people use
+the Tor network on a daily basis.
+Counting users in an anonymity network is, obviously, a difficult task
+for which we cannot collect too sensitive usage data.
+We came up with a privacy-preserving approach for estimating directly
+connecting user numbers by counting requests to the directory mirrors and
+deriving approximate user numbers from there.
+In this report we describe a modified approach for estimating the number
+of users connecting via bridges by evaluating directory requests made to
+bridges.
+We compare this new approach to our current approach that estimates bridge
+user numbers from total unique IP addresses seen at bridges.
+We think that results from the new approach are closer to reality, even
+though that means there are significantly fewer daily bridge users than
+originally expected.
+\end{abstract}
+
+\section{Introduction to our new approach to count bridge users}
+
+In this report we describe a new approach for estimating the number of
+daily users connecting to the Tor network via a bridge.
+This new approach uses counts of directory requests made to bridges as its
+main data sources.
+This is similar to how we estimate daily directly connecting users that
+connect to the Tor network via a non-bridge relay.
+Our current approach for estimating daily bridge users is to count unique
+IP addresses of connecting clients at bridges.
+We refer to our earlier report~\cite{tor-2010-11-001} for an overview of
+estimating user numbers in the Tor network.
+
+We estimate daily bridge users by first summing up directory requests per
+day reported by bridges (Section~\ref{sec:reports}).
+We extrapolate these reported requests to the expected total number of
+directory requests in the network (Section~\ref{sec:requests}).
+We then assume that there is an average number of 10 directory requests
+that every client makes per day and derive daily user numbers by dividing
+by that average number (Section~\ref{sec:users}).
+We further derive users per country by including country information of
+connecting IP addresses (Section~\ref{sec:country}).
+There are at least two ways to remove unwanted artifacts from results:
+we may have to ignore reports from bridges that have been running as
+non-bridge relays and that might still report directly connecting users
+(Section~\ref{ref:nonbridge});
+and we may need to ignore days when there were problems with the consensus
+process, leading to an increase in directory requests which is likely not
+caused by an actual increase in users (Section~\ref{ref:consensus}).
+
+\section{Counting reported directory requests to bridges per day}
+\label{sec:reports}
+
+All relays running a recent enough Tor version contain code to collect and
+report statistics on processed directory requests over 24-hour periods.
+We refer to our previous work~\cite{loesing2010case} for more details on
+aggregating usage statistics in Tor.
+
+Bridges were originally not supposed to report directory request
+statistics, because bridge users were estimated based on unique IP address
+counts, and we wanted to avoid collecting any more data than needed.
+But due to a bug, only part of directory request statistics were
+suppressed when running in bridge relay mode.
+As a result, bridges report a trimmed version of directory request
+statistics, which are however still enough to estimate daily users.
+As an example, the following directory request statistics were reported by
+a bridge in September 2012.
+
+%All examples come from descriptor FE2E5A0E7B15F5C1E102791BC957A6F9BE9DDC2E
+\begin{quotation}\noindent\footnotesize
+\texttt{%
+extra-info goinpostal 7363FF835F5D79EA1F0CC2EB757B03866D4515F7\\
+dirreq-stats-end 2012-09-18 15:26:38 (86400 s)\\
+dirreq-v3-resp ok=5040,not-enough-sigs=0,unavailable=0,not-found=0,not-modified=0,busy=0}
+\end{quotation}
+
+From this example we learn that this bridge successfully processed 5,040
+version 3 directory requests in the 24 hours preceding September 18, 2012,
+15:26:38 UTC.
+We have no information how many of these requests happened in the 8.5
+hours of September 17 or in the 15.5 hours of September 18.
+We assume a uniform distribution of requests over the 24-hour interval and
+count 1,797 requests for September 17 and 3,243 requests for September 18.
+We extract these data points for all bridges publishing their descriptors
+to the bridge authority and sum up responses per day.
+Figure~\ref{fig:responses} contains the number of reported requests per
+day.
+\begin{figure}
+\includegraphics[width=\textwidth]{responses.pdf}
+\caption{Daily sums of directory requests reported by all bridges}
+\label{fig:responses}
+\end{figure}
+A few observations:
+
+\label{lab:discuss-responses}
+\begin{enumerate}
+\item There were hardly any reported requests until September 2011.
+The likely explanation is that, before Tor version 0.2.3.1-alpha,
+collecting and reporting directory request statistics was disabled by
+default.
+This default was changed in Tor version 0.2.3.1-alpha which was released
+on May 5, 2011.
+With more and more bridges upgrading to the 0.2.3 series, the number of
+reported directory requests increased, too.
+\item The reported request numbers in September and October 2011 have
+quite high volatility, making it difficult to use these request numbers
+for actual user number estimates.
+\item There are at least three unusual spikes in request numbers in
+November 2011, January 2012, and April 2012, which were unlikely caused by
+sudden increases and decreases in user numbers.
+\item From November 2011 on, there is a general downward trend from around
+60,000 requests per day to just 30,000 in September 2012.
+\end{enumerate}
+
+\section{Extrapolating to total directory requests in the network}
+\label{sec:requests}
+
+So far, we only know how many directory requests were processed by the
+bridges reporting them.
+We need to take into account that not all bridges report these statistics
+for various reasons:
+bridges may not be configured to report directory request statistics,
+which in particular applies to bridges running an earlier version than
+0.2.3.1-alpha;
+bridges may run for less than 24 hours, thus not finishing a 24 hour
+statistics interval and discarding requests processed up to that time;
+bridges may have finished a 24-hour statistics interval, but went offline
+before publishing statistics to the bridge authority.
+(We analyzed in more detail what fraction of our bridges are not reporting
+usage statistics in~\cite{tor-2012-04-001}.)
+As a result, we need to extrapolate reported requests to what we expect as
+the total number of requests in the network.
+
+A straight-forward way to extrapolate to the total number of directory
+requests in the network would be to make the following assumption:
+every bridge that does not report directory request statistics, on
+average, processes as many directory requests as a bridge that reports
+them.
+Under this assumption, we could count the number of running bridges per
+day and the number of bridges reporting directory request statistics,
+compute the fraction of reporting bridges, and divide reported requests by
+that fraction.
+This assumption works fine as long as the variance between request numbers
+processed by bridges is small, or as long as the fraction of reporting
+bridges is high.
+However, the former is not the case, because there are some hard-coded
+bridge addresses in bundles distributed on the Tor website, leading to
+these bridges processing and maybe reporting far more directory requests
+than others.
+The latter is not always the case either, in particular before September
+2011, as we could see in Figure~\ref{fig:responses}.
+
+A better way to extrapolate to total requests in the network is to
+consider a second statistic published by a subset of bridges: the number
+of bytes written to respond to directory requests.
+We assume that this number is proportional to the number of processed
+directory requests, even though we do not assume an exact linear relation.
+The subset of bridges reporting byte statistics is not necessarily the
+same subset thas is reporting directory request statistics.
+By taking into account byte histories, we can better estimate how many
+directory requests have not been reported by bridges that at least have
+reported byte histories.
+An example for reported written directory bytes, coming from the same
+extra-info descriptor as the example above, is as follows:
+
+\begin{quotation}\noindent\footnotesize
+\texttt{%
+extra-info goinpostal 7363FF835F5D79EA1F0CC2EB757B03866D4515F7\\
+dirreq-write-history 2012-09-19 05:17:32 (900 s) 13370368,10539008,56751104,27235328,\\
+14555136,10524672,63341568,37339136,24343552,22490112,29155328,17792000,3502080,[...]}
+\end{quotation}
+
+From these write histories we can extract how many bytes a bridge has
+spent on answering directory requests on a given day.
+By looking at the reported directory request history values, we can learn
+how many bytes were written while the bridge was also collecting and later
+reporting directory request statistics, and we can learn how many bytes
+were written outside of those statistics intervals.
+Similarly, we can learn how many directory requests were processed at
+times when the bridge did not report byte histories.
+
+On a side note, as one can see from the example, byte history intervals
+are only 15 minutes long and thereby much shorter than directory request
+statistics intervals.
+It might be that this level of detail has privacy implications, in
+particular on bridges with only very few users.
+We probably don't need byte histories on this level of detail.
+We leave the analysis what level of detail is required as future work.
+Results could lead to increasing the history interval length on bridges to
+1 hour or more.
+
+In the following, we define $R$ as the subset of bridges reporting
+directory request statistics, $H$ as the subset reporting byte histories,
+and $N$ as the entire set of bridges in the network.
+Figure~\ref{fig:bridgesets} illustrates these subsets and the variable
+names.
+Also, we define $r()$ as the number of directory requests reported by a
+given set of bridges, $h()$ as the number of bytes reported by the bridges
+in a given set, and $n()$ as the absolute number of bridges in a set.
+
+\begin{figure}
+\centering
+\includegraphics[width=.5\textwidth]{bridgesets.pdf}
+\caption{Subsets of bridges reporting write histories and directory
+request statistics}
+\label{fig:bridgesets}
+\end{figure}
+
+Knowing the number of reported directory requests, $r(R)$, we can
+extrapolate to the expected total number of directory requests, $r(N)$, by
+multiplying with the reciprocal of the fraction of written directory
+requests that got reported to us, $\frac{h(N)}{h(R)}$:
+
+\begin{equation}
+r(N) = r(R) \times \frac{h(N)}{h(R)}
+\end{equation}
+
+Estimating total written directory request bytes in the network, $h(N)$,
+is easy.
+We assume here that the bridges that didn't report directory request bytes
+wrote the same number of bytes per bridge on average as reporting bridges.
+
+\begin{equation}
+h(N) = h(H) \times \frac{n(N)}{n(H)}
+\label{equation}
+\end{equation}
+
+Estimating written directory request bytes by the bridges that reported
+directory request statistics, $h(R)$, is somewhat harder.
+We first split the set into the set of bridges reporting both statistics
+and the set of bridges reporting only statistics and no write histories.
+
+\begin{equation}
+h(R) = {h(R \cap H) + h(R \setminus H)}
+\end{equation}
+
+The first number is something we can read from the descriptors.
+The second number requires us to apply the same assumption from above,
+namely that bridges that didn't report byte histories wrote the same
+number of bytes, on average, as reporting bridges.
+(Note how this equation is very similar to equation \ref{equation}.)
+
+\begin{equation}
+h(R \setminus H) = h(H) \times \frac{n(R \setminus H)}{n(H)}
+\end{equation}
+
+Putting everything together, we come up with a way to compute estimated
+directory requests in the network:
+
+\begin{equation}
+r(N) = r(R) \times \frac{h(H) \times n(N)}{h(R \cap H) \times n(H) + h(H) \times n(R \setminus H)}
+\end{equation}
+
+Figure~\ref{fig:totalrequests}
+\begin{figure}[t]
+\includegraphics[width=\textwidth]{extrapolated-responses.pdf}
+\caption{Reported directory requests, estimated fraction of bridges
+reporting directory requests, and estimated directory requests in the
+network}
+\label{fig:totalrequests}
+\end{figure}
+shows reported directory requests, estimated
+fraction of directory requests that got reported by bridges, and estimated
+total directory requests in the network.
+A few observations:
+
+\begin{enumerate}
+\item The top-most graph is the same as in Figure~\ref{fig:responses} which
+we already discussed on page~\pageref{lab:discuss-responses}.
+\item The middle graph shows an upwards trend of the fraction of bridges
+reporting directory request statistics.
+Fractions of under 25\% as seen until end of October 2011 make it
+difficult to extrapolate to the total number of requests in the network.
+These fractions also explain the observed volatility of reported requests
+until end of October 2011.
+Beginning with November 2011, fractions are at 50\% or higher, exceeding
+75\% in most of 2012.
+\item The bottom-most graph is the result of dividing request numbers in
+the top-most graph by fractions in the middle graph.
+The low fractions in late August and early September 2011 lead to very
+high estimated total requests in the network.
+We'll want to treat these surprisingly high numbers with care, but so far,
+there is no reason to believe they're totally wrong.
+From November 2011 to September 2012, the continuously decreasing reported
+request numbers combined with the increasing fraction of reported requests
+lead to a continuous decrease in estimated directory requests in the
+network.
+From this graph it seems that bridge usage has steadily decreased in the
+past 11 months.
+\end{enumerate}
+
+\section{Dividing by 10 for estimating number of users}
+\label{sec:users}
+
+With the estimated number of daily directory requests in the network we
+can now estimate the number of daily users.
+We make the assumption that there is an average number of directory
+requests per day that every client makes to keep their network information
+up-to-date.
+As of writing this report, network status consensuses are fresh for three
+hours, requiring clients to download a new document every 2 to 3 hours.
+Hence, a client that is online all day would make 8 to 12 directory
+requests on that day.
+Not all clients are online all day, thus reducing the average number of
+directory requests made by clients.
+We, somewhat arbitrarily, chose 10 as the number of directory requests
+that the average client makes every day.
+10 is also the number that we use in directly connecting user statistics,
+so that both estimates for non-censored and censored users will be easy
+to compare.
+We could evaluate whether 10 is a good number by asking volunteers to have
+their Tor clients record directory request numbers made on a given day,
+and use these actual numbers to come up with a better number.
+But given that we apply the same number of requests per client to all
+days, the actual value does not influence development over time, allowing
+us to observe trends over time and still have a rough idea of absolute
+numbers.
+
+Now we can derive user numbers from total directory requests in the
+network.
+Figures~\ref{fig:totalusers} and \ref{fig:totalusers-q3-2012}
+show the total number of users connecting via bridges over time, for the
+entire period for which we have data and for the third quarter of 2012.
+These graphs contain the same data as the bottom-most graph in
+Figure~\ref{fig:totalrequests}, but divided by 10.
+The most surprising result is that there are only 3,500 daily bridge users
+these days.
+
+\begin{figure}
+\includegraphics[width=\textwidth]{totalusers.pdf}
+\caption{Estimated daily bridge users from all countries from July 2011 to
+September 2012}
+\label{fig:totalusers}
+\end{figure}
+\begin{figure}
+\includegraphics[width=\textwidth]{totalusers-q3-2012.pdf}
+\caption{Estimated daily bridge users from all countries in the third
+quarter of 2012}
+\label{fig:totalusers-q3-2012}
+\end{figure}
+
+\section{Breaking down to user numbers by country}
+\label{sec:country}
+
+So far, we only have an estimate of daily bridge users in the network, but
+no user numbers per country.
+In contrast to directory request statistics reported by relays, bridges
+do not report request numbers by country code.
+But we still have statistics on originating countries in the unique IP
+address statistics reported by bridges.
+These are the statistics we have been using for estimating daily bridge
+users so far.
+We assume that the country distribution of connecting bridge clients is
+similar to bridge clients downloading directory requests.
+As an example, these are the statistics on connecting clients reported by
+the same bridge from earlier examples:
+
+\begin{quotation}\noindent\footnotesize
+\texttt{%
+extra-info goinpostal 7363FF835F5D79EA1F0CC2EB757B03866D4515F7\\
+bridge-stats-end 2012-09-18 15:27:00 (86400 s)\\
+bridge-ips ir=32,??=16,at=8,eg=8,jp=8,lv=8,pk=8,us=8}
+\end{quotation}
+
+We sum up unique IP addresses and calculate a fraction of IP addresses for
+every country and day.
+(We could also have weighted country information of reporting bridges with
+the bridge's fraction of written directory request bytes, but this seemed
+like overkill for this analysis, so we left it for future work.)
+We multiply the estimated number of total users in the network with the
+fraction of unique IP addresses coming from a country and come up with the
+estimated number of users in that country.
+Figure~\ref{fig:syusers.pdf} shows the result for estimated daily bridge
+users coming from Syria.
+We chose Syria for this example, because that's one of the countries with
+most bridge users these days.
+The approach would work for all other countries, too.
+
+\begin{figure}
+\includegraphics[width=\textwidth]{syusers.pdf}
+\caption{Estimated daily bridge users from Syria}
+\label{fig:syusers.pdf}
+\end{figure}
+\section{Ignoring bridges that were running as non-bridge relays}
+
+\label{ref:nonbridge}
+
+One aspect that we have been ignoring so far is that there are bridges
+that were running as non-bridge relays before.
+It is likely that these bridges report directory requests coming from
+directly connecting clients of which there are far more than bridge
+clients.
+
+Figure~\ref{fig:responses-single-bridges} shows reported directory
+requests of bridges that have or have not been seen as non-bridge relays.
+The focus here is on data points which are seemingly outliers.
+Two noteworthy examples are the 30,000+ requests in April 2012 and the
+25,000+ requests in September 2012.
+
+In our current approach to count daily bridge users, we ignore any such
+bridge, because they could skew results.
+We'd like to exclude data points coming from bridges that report
+unrealistic statistics.
+However, it seems that ignoring all bridges that have been seen as
+non-bridge relays would mean removing too many data points.
+More research is needed to define criteria when a data point probably
+contains directory requests by non-bridge clients and should be ignored.
+
+\begin{figure}
+\includegraphics[width=\textwidth]{responses-single-bridges.png}
+\caption{Reported directory requests by bridges that have or have not been
+seen as non-bridge relays}
+\label{fig:responses-single-bridges}
+\end{figure}
+
+\section{Ignoring days with too few network status consensuses}
+\label{ref:consensus}
+
+A closer look at spikes in total estimated directory requests and at
+archives of network status consensuses reveals an interesting correlation:
+in 4 out of 5 cases when less than 20 consensuses were published on a given
+day, the number of directory requests went up a lot.
+Figure~\ref{fig:consensuses} shows the number of published consensuses per
+day.
+
+\begin{figure}
+\includegraphics[width=\textwidth]{consensuses.pdf}
+\caption{Published consensuses per day}
+\label{fig:consensuses}
+\end{figure}
+
+By default, the directory authorities publish a new consensus every hour.
+Missing consensuses indicate a problem with the consensus process, meaning
+that the available consensuses become outdated and that clients send out
+more requests to get a recent enough consensus, thus raising the number of
+directory requests in the network.
+
+We could decline providing bridge usage statistics for days when the
+archives had less than 20 consensuses.
+This would fix the spike in late November 2011 and especially the huge one
+in early January 2012.
+However, we decided not to remove these days yet and left it as future
+work to analyze how we can detect problems with the consensus process
+leading to higher directory request numbers.
+
+\section{Comparing old and new approaches to count bridge users}
+
+We briefly compare results from our new approach based on directory
+requests to results from our existing approach based on unique IP
+addresses.
+Figures~\ref{fig:compare-totalusers} and
+\ref{fig:compare-totalusers-q3-2012} show the estimates of daily bridge
+users in the new and in the old approach, for the entire observed period
+and for the third quarter of 2012.
+The general trend is about the same, though the new
+approach only outputs about one tenth as many daily bridge users as the old
+approach did.
+We think that results from the new approach are closer to reality, because
+the reasoning behind it is much more plausible than the design of the old
+approach.
+We refer to our earlier report~\cite{tor-2010-11-001} for details of the
+old approach to count daily bridge users including a discussion of its
+weaknesses.
+
+\begin{figure}[t]
+\includegraphics[width=\textwidth]{compare-totalusers.pdf}
+\caption{Estimated daily bridge users in the new and in the old approach
+from July 2011 to September 2012}
+\label{fig:compare-totalusers}
+\end{figure}
+
+\begin{figure}[t]
+\includegraphics[width=\textwidth]{compare-totalusers-q3-2012.pdf}
+\caption{Estimated daily bridge users in the new and in the old approach
+in the third quarter of 2012}
+\label{fig:compare-totalusers-q3-2012}
+\end{figure}
+
+\section{Suggesting next steps}
+
+We identified a few starting points for further improving the described
+approach in this report:
+
+\begin{itemize}
+\item We should re-run the analysis under the assumption that bridges use
+a larger time period for byte histories than 15 minutes.
+There are potential privacy problems with seldomly used bridges reporting
+usage statistics on such a high detail level.
+The described approach to estimate daily users should work as well with
+byte histories on a detail of a few hours.
+Once we know what detail is required, we should change the default in the
+Tor sources, and we could update the bridge descriptor sanitizing code to
+increase the byte history interval in sanitized bridge descriptors.
+\item We should evaluate whether 10 is a reasonable average number of
+directory requests made by a client per day.
+One way to do this evaluation is to ask volunteers to have their Tor
+clients record directory request numbers made on a given day.
+\item We should look closer at weighting country information reported by
+bridges.
+Maybe we'll have to weight unique IP address fractions by country with the
+reporting bridge's written directory request byte fraction to get more
+accurate user numbers by country.
+\item We should further investigate how bridges that have been seen as
+non-bridge relays affect the results.
+If we need to ignore reported statistics by these bridges, we'll want to
+make sure to only exclude as few reports as necessary.
+\item We should further analyze problems with the consensus process and
+how they affect directory request numbers.
+Both statistics of daily directly connecting and of daily bridge users
+could benefit from new insights here.
+\end{itemize}
+
+\section*{Acknowledgements}
+
+Thanks to George Kadianakis for his valuable input on a draft of this
+report, especially by reviewing the math part of extrapolating reported
+requests to total requests in the network, and for pointing out the
+potential privacy problem of too short byte history intervals.
+
+\bibliography{counting-daily-bridge-users}
+
+\end{document}
+
diff --git a/2012/counting-daily-bridge-users/extrapolated-responses.pdf b/2012/counting-daily-bridge-users/extrapolated-responses.pdf
new file mode 100644
index 0000000..24d3fb1
Binary files /dev/null and b/2012/counting-daily-bridge-users/extrapolated-responses.pdf differ
diff --git a/2012/counting-daily-bridge-users/responses-single-bridges.png b/2012/counting-daily-bridge-users/responses-single-bridges.png
new file mode 100644
index 0000000..199be4e
Binary files /dev/null and b/2012/counting-daily-bridge-users/responses-single-bridges.png differ
diff --git a/2012/counting-daily-bridge-users/responses.pdf b/2012/counting-daily-bridge-users/responses.pdf
new file mode 100644
index 0000000..f974b19
Binary files /dev/null and b/2012/counting-daily-bridge-users/responses.pdf differ
diff --git a/2012/counting-daily-bridge-users/syusers.pdf b/2012/counting-daily-bridge-users/syusers.pdf
new file mode 100644
index 0000000..ff52022
Binary files /dev/null and b/2012/counting-daily-bridge-users/syusers.pdf differ
diff --git a/2012/counting-daily-bridge-users/tortechrep.cls b/2012/counting-daily-bridge-users/tortechrep.cls
new file mode 120000
index 0000000..4c24db2
--- /dev/null
+++ b/2012/counting-daily-bridge-users/tortechrep.cls
@@ -0,0 +1 @@
+../../tortechrep.cls
\ No newline at end of file
diff --git a/2012/counting-daily-bridge-users/totalusers-q3-2012.pdf b/2012/counting-daily-bridge-users/totalusers-q3-2012.pdf
new file mode 100644
index 0000000..a2d27d5
Binary files /dev/null and b/2012/counting-daily-bridge-users/totalusers-q3-2012.pdf differ
diff --git a/2012/counting-daily-bridge-users/totalusers.pdf b/2012/counting-daily-bridge-users/totalusers.pdf
new file mode 100644
index 0000000..1a982ef
Binary files /dev/null and b/2012/counting-daily-bridge-users/totalusers.pdf differ
More information about the tor-commits
mailing list