An Empirical Analysis of Web Censorship Lists

class: center, middle

# Topics of Controversy
## <a href="https://research.owlfolio.org/pubs/2017-topics-controversy.pdf">An Empirical Analysis of Web Censorship Lists</a>

.authors[
<a href="https://www.owlfolio.org/">Zachary Weinberg</a>,
Mahmood Sharif, 
Janos Szurdi,
<a href="https://www.cylab.cmu.edu/education/faculty/christin.html">Nicolas Christin</a>
]

invisible italic
 and bold text to force fonts to load

???
Hello everyone.  I’m Zachary Weinberg, and I’m presenting my paper
“Topics of Controversy, an empirical analysis of web censorship
lists.”  This is joint work with Mahmood Sharif, Janos Szurdi, and
Nicolas Christin, we’re all at Carnegie Mellon.

---
# How Everyone Measures Web Censorship

1. Make a list of webpages that might be censored
1. Get a “vantage point” computer in the country of interest
1. Try to load all the webpages from step 1
1. Did we get suspicious errors or block screens?

This talk is about step 1: how do you make a good list?

???
[recite the four steps]

You could write an entire paper about each of these steps, and this is
the paper about step 1.  How do you know if you have made a good list?

---
# Criteria for Probe Lists

* _Breadth:_ many different types of potentially censored material
* _Depth:_ many sites for each type of material
* _Freshness:_ active sites, not abandoned ones
* _Efficiency:_ list can be probed quickly w/ slow network
* _Maintainability:_ little manual work to keep up to date

This talk is about _breadth_ and _freshness._

???
Here’s some criteria.  Breadth, depth, freshness, efficiency,
maintainability.  And again you could probably write a whole paper
about each one.  This paper tackles breadth and freshness.

---
# Some existing lists

???
Here are some existing lists. 22 of them, 760,000 unique URLs. Three
broad categories.

15 alleged actual blacklists. One-time snapshots, easy to
over-interpret, some from low-credibility sources.  The top eight
(“pinklists”) have overwhelmingly more porn than anything else, you
can tell just from looking at the URLs.  The other seven
(“blacklists”) have more variety.

And seven comparison lists, selected seven different ways.  Three
should have more censored pages than average; the “probe list” was
created by the OpenNet Initiative specifically to probe censorship.
One should have _fewer_ censored pages than average, and the final
three are samples of the more popular parts of the Web.  Common Crawl
in particular is intended to be a completely uniform sample, and we
use it as the “average” baseline in our analyses.

---
# Some existing lists compared

???
This matrix shows Jaccard similarity — what proportion of the URLs in
list 1 also appear in list 2, and vice versa?

We can see that the pinklists have a lot in common with each other,
the comparison lists also have a lot in common, and most of the
blacklists _don’t_ have a lot in common with either other group, or
with each other.  Not very helpful.  How can we make a more meaningful
comparison?

Our key hypothesis: the reason a page is censored will usually have
something to do with its _topic_.  So comparing the topic distribution of
each list should tell us something interesting.

---
# Determining Topics of Webpages

* Want _uncensored_ copies of as many pages as possible
  * collect from a data center in the USA
  * use a headless browser for accuracy
* Pages may have been taken down or changed
  * exclude empty pages and HTTP/DNS/network errors (42%)
  * exclude parked domains (additional 7%)
  * backfill from the Internet Archive when possible
* Topic analysis is AI-hard if you want to do it _right_
  * language, imagery, idiom, context, discourse, culture, history
  * Crude approximations work surprisingly well

???

We downloaded _uncensored_ copies of all the pages on the lists —
assuming that a data center in New Jersey sees a totally uncensored
Web — and did some NLP on them to identify their topics.

Of course, just downloading 760 thousand pages is nontrivial.  We have
to use a headless browser and run JavaScript, and nearly half of the
pages have been taken down already so we have to detect that and
backfill from the Internet Archive if we can.

And then, once you have 760 thousand pages, programming a computer to
understand “what they are about” _properly_ would involve creating a
human-equivalent artificial intelligence.  Which would then refuse to
do the job.

But the great thing about AI-hard problems is that crude
approximations are often plenty good enough.  I’m going to walk you
through our process for one webpage.

---
# Example webpage, as downloaded

???
It’s got a fancy banner and some pictures and navigation links and
maybe an ad off at the side.  We don’t want any of that.  (We don’t
even download the images.)

---
# Reduced to the main text

???
We extract the main text using heuristics, based on the principle
that more heavily marked-up text is likely to be boilerplate.

---
# Languages identified

???
Forty percent of our documents are in multiple languages.  Here we’ve
got some Latin and some Greek.

---
# Translated word by word into English

???
We couldn’t afford to feed entire documents into Google Translate so
we went word by word instead.  We have to have everything in the same
language for the next step.

---
# LDA-clustered with similar documents

???
LDA, Latent Dirichlet Analysis, is a standard technique for clustering
“bags of words.”  It gives us clusters of documents that have more in
common with each other than with the rest of the database, and for
each cluster, a list of its characteristic words.

The characteristic words usually make sense to a human, you can stick
a one- or two-word label on them...

---
# Cluster manually labeled

???
... for instance, in this case, Stoic philosophy.

---
# 64 topics, 9 categories

<table class="longlist"><tr>
<td style="width:100px"><ul>
<li>news and politics<ul>
 <li>general</li>
 <li>world</li>
 <li>business</li>
 <li>legal</li>
 <li>environment</li>
 <li>human rights</li>
 <li>extremism</li>
 <li>war</li>
 <li>USA</li>
 <li>Europe</li>
 <li>France</li>
 <li>Germany</li>
 <li>Italy</li>
 <li>Spain</li>
 <li>Brazil</li>
 <li>Russia</li>
 <li>Middle East</li>
 <li>Israel</li>
 <li>Iran</li>
 <li>India</li>
 <li>East Asia</li>
 <li>China</li>
 <li>Japan</li>
 <li>Thailand</li>
 <li>Vietnam</li>
 <li>Oceania</li>
</ul></li>
</ul></td>
<td style="width:100px"><ul>
<li>commerce<ul>
 <li>online shopping</li>
 <li>fashion</li>
 <li>marketing</li>
 <li>manufacturing</li>
 <li>electronics</li>
</ul></li>
<li>religion<ul>
 <li>general</li>
 <li>Christianity</li>
 <li>Islam</li>
</ul></li>
<li>scholarship<ul>
 <li>science</li>
 <li>medicine</li>
 <li>history</li>
 <li>education</li>
 <li>libraries</li>
 <li>books</li>
</ul></li>
</ul></td>
<td style="width:100px"><ul>
<li>entertainment<ul>
 <li>art</li>
 <li>travel</li>
 <li>wine</li>
 <li>food</li>
 <li>sports</li>
 <li>games</li>
 <li>music</li>
 <li>celebrities</li>
 <li>social media</li>
 <li>gambling</li>
 <li>dating</li>
</ul></li>
<li>video (not porn)<ul>
 <li>general</li>
 <li>television</li>
</ul></li>
</ul></td>
<td style="width:125px"><ul>
<li>porn<ul>
 <li>general</li>
 <li>kinky</li>
 <li>gay</li>
 <li>age gates</li>
</ul></li>
<li>software<ul>
 <li>sales</li>
 <li>development</li>
 <li>servers and hosting</li>
</ul></li>
<li>junk<ul>
 <li>web page boilerplate</li>
 <li>error messages</li>
 <li>mistranslated</li>
 <li>unintelligible</li>
 <li>empty after boilerplate removal</li>
 <li>unable to translate document</li>
</ul></li>
</ul></td></tr></table>

???
Here’s the labels for all of the pages we downloaded.

Despite the work we put into cleaning up the documents before
clustering, we didn’t get rid of all the boilerplate, or even all the
errors — some major, famous websites serve error messages with HTTP
status 200 — don’t do that, ok? and we couldn’t translate everything,
either.

But the overall set of topics generally makes sense and doesn’t have
any obvious gaps.

---
# Topics and Lists

.center[<img src="source-topic-all.svg" class="fullwidth">]

???
This is a correlation matrix between the source lists and the topics.
Very briefly, whereever you see red, that means some list has more of
some topic than average; blue means _less_ of some topic than average.
(“Average” is defined by the Common Crawl sample at the very bottom,
that’s why it’s all white.) Blue stripes don’t mean much, but vertical
_red_ stripes tell us that some group of lists share a bias toward a
particular topic.

There’s a lot going on here, so let’s zoom in on a couple of
interesting cases.

---
# Blacklists and ONI’s probe list

.center[<img src="source-topic-sub1.svg" class="fullwidth">]

???

The OpenNet Initiative is primarily concerned with political
censorship, so their probe list includes a wide variety of
news-and-politics topics.  So do the blacklists — but they tend to be
more concentrated on _specific_ news topics.  India, Russia, Syria,
Thailand are all censoring news related to _themselves_ and their
immediate neighbors, and they don’t care so much about other
countries.

Over on the far right, we can also see that ONI may be underestimating
the amount of censorship related to social media.  There’s lots of
evidence that authoritarian governments treat social media as a
primary threat, so that’s unfortunate.

We also see more entertainment topics—music, video, TV—in the
blacklists than ONI anticipates.  It seems likely that censorship
infrastructure is being reused to go after copyright infringement.  Or
maybe that’s what it was built for in the first place.

---
# The pinklists aren’t all porn

.center[<img src="source-topic-sub2.svg">]

???

The classifier identifies a number of sites on the German pinklist as
relating to German and extremist politics.  I spot-checked these and,
yep, neo-Nazism.  (It’s not always obvious — there’s a site in there
that sells sheet music of Central European folk songs.)

Similarly, the classifier finds some Thai and Chinese political sites
and social media in the 2007 Thailand pinklist, and they’re
consistently critical of the Thai royal family.  This is known, overt
policy in both cases, but I didn’t have to wade through all the porn
to find it.

On the other hand, the Russia-related sites on Australia, Denmark,
Italy, and Norway’s pinklists are a misclassification: spot-checking
suggests that a lot of those are sketchy dating websites.

---
# Site Lifetimes

.center[<img src="km-topic-cat.svg" style="width:85%;margin-top:-15px">]

???

Now let’s talk about freshness.  Many of the lists, especially the
pink and black lists, are five to ten years old.  Overall, half of the
pages are gone already.  We used the Internet Archive to find out how
long ago each page got taken down, and how long it lasted.

This is a Kaplan-Meier survival estimate for each category of topics.
It shows the probability that a page is still up after some number of
years since it was first observed by the Archive.

Regardless of topic, roughly a quarter of the pages are gone after
four years, and then there’s a bit of a knee.  Pages classified as
religion, commerce, or scholarship tend to live longer than most.

---
# Odds of Page Removal

.center[<img src="hazardmodel.svg" class="fullwidth">]

???

But what we really want to know is, as censorship researchers, do we
need to worry more or less than average about pages going away?  This
is a Cox “proportional hazards” model, which predicts the odds of a
page being taken down as a function of its topic category and its
list category, relative to an average (Common Crawl again).

What it says is, yes we do need to worry more.  The fat blue line in
the middle is the average.  Unfortunately, _everything_ is above
average except non-pornographic video.  Youtube has been around for a
long time and isn’t going anywhere.

What’s really important is that each letter — topic category — in the
right three panels — probably censored, blacklist, pinklist — is
higher up the chart than the same letters in the left two panels.  The
vertical axis is logarithmic, so we’re talking five to ten times more
likely to go away, here.

And note that the non-pornographic stuff hiding in the pinklists is
more likely to go away than anything else.

---
# Conclusions

* Censored topics are highly country-specific
* Censored sites are more likely to go away than average
* Topic clustering can reveal details of “leaked” censorship lists
* Topic clustering should help us maintain probe lists