[00:00.000 --> 00:00.700]  Jonathan Mair.
[00:04.840 --> 00:10.480]  Before getting into the substance, I want to make sure to note that everything I'm about to present is not just my work.
[00:10.640 --> 00:13.220]  I had some awesome collaborators on all of this.
[00:13.980 --> 00:23.860]  And Dan Benet and John Mitchell, faculty at Stanford, Ed Felton, faculty at Princeton, my colleague Patrick at Stanford Computer Science had tremendous input.
[00:24.040 --> 00:25.500]  None of this is just my own work.
[00:26.620 --> 00:30.640]  I also have to do my obligatory disclaimer at this point, that I'm not your lawyer.
[00:31.500 --> 00:33.760]  And nothing I'm going to say today is legal advice.
[00:34.540 --> 00:41.880]  And if you have questions you want to ask that are of a legal nature, this isn't a confidential environment, so keep that in mind.
[00:42.400 --> 00:47.140]  And the reason you know I'm not your lawyer and this isn't legal advice is because the transaction would look a lot more like that.
[00:49.200 --> 00:56.740]  Okay, so let me start with the motivation for our research on the NSA.
[00:57.780 --> 01:05.540]  And it started from a place that I think many of you may be familiar with, a lot of frustration with the agency.
[01:05.540 --> 01:09.880]  That it hadn't been very straightforward, that it engaged in a lot of conduct that we found objectionable.
[01:11.080 --> 01:14.920]  And so, you know, we were pretty concerned about what we were learning.
[01:15.500 --> 01:18.340]  And there was sort of a common mode of response.
[01:18.360 --> 01:20.040]  Again, I'm sure many of you are familiar with this.
[01:22.140 --> 01:25.820]  That the NSA deserved some pretty substantial public opprobrium.
[01:28.280 --> 01:28.840]  Yeah.
[01:28.840 --> 01:36.220]  So that, you know, seemed to us an entirely fair viewpoint for folks to have.
[01:36.760 --> 01:48.320]  But it seemed to us maybe there was a way of leveraging computer science and debates over the NSA to try to have a constructive role alongside the sort of directed opprobrium at the agency.
[01:48.320 --> 01:56.120]  And so kind of very roughly, this was our model of how the NSA interacted with its oversight authorities.
[01:57.300 --> 02:08.080]  And maybe it would make sense if we could find some way to get some actual computer science facts into this debate in a way that had seemed, unfortunately, quite lacking.
[02:09.120 --> 02:14.480]  And so we started by trying to learn as much as we could about the law that governs the NSA.
[02:15.380 --> 02:18.660]  Talking to intelligence law experts all over the place.
[02:19.080 --> 02:23.720]  Then did some computer science research, trying to understand the implications of the law.
[02:23.960 --> 02:39.240]  And then finally, returned that to legal fora for arguments about why, as a matter of what we now know from computer science, the NSA's program should be held unconstitutional or should make less sense as a matter of policy from one of the oversight board's perspectives.
[02:39.820 --> 02:41.740]  So this is what we tried to do.
[02:43.180 --> 02:51.480]  And thinking through the law, it became very clear to us early on that domestic bulk surveillance authorities include privacy safeguards.
[02:52.360 --> 02:56.160]  They're safeguards primarily for U.S. persons.
[02:56.200 --> 03:00.220]  So if you're a foreigner outside the United States, you're kind of out of luck.
[03:02.280 --> 03:05.620]  And today I'm not going to be talking about international surveillance.
[03:05.620 --> 03:11.240]  In particular, you might have heard of Executive Order 12333 and what the NSA does outside the United States.
[03:11.420 --> 03:13.980]  I'm not planning to talk about that in the body of today's talk.
[03:14.100 --> 03:15.420]  I'm happy to take questions about it.
[03:15.580 --> 03:20.040]  I'm also not going to be talking about individualized surveillance.
[03:20.580 --> 03:23.860]  Targeted orders for an individual's information.
[03:24.080 --> 03:26.700]  Since the legal structures there look very different.
[03:26.700 --> 03:32.480]  And I think it's fair to say raise a very different sort of policy concern from bulk surveillance authorities.
[03:33.940 --> 03:40.860]  So when thinking through these privacy safeguards, it became apparent to us quite early that these were safeguards designed by lawyers.
[03:40.880 --> 03:46.820]  And that was kind of a weird property for them since there are safeguards on primarily technical systems.
[03:46.820 --> 03:56.640]  So lawyers on the Supreme Court, the Foreign Intelligence Surveillance Court, certainly within the executive branch, had put together a set of rules.
[03:57.180 --> 03:59.760]  And rules based on technical assumptions.
[04:01.120 --> 04:04.680]  Maybe it would make sense to like, you know, if you can't practice law without a license.
[04:04.680 --> 04:08.400]  Maybe you can't practice like tech law without a computer science license or something like that.
[04:09.520 --> 04:13.480]  But anyway, so these guys were kind of practicing computer science without a license.
[04:13.480 --> 04:16.380]  And so we thought, okay, let's take a look at those assumptions.
[04:17.080 --> 04:18.800]  Let's see if they're accurate.
[04:19.560 --> 04:23.760]  And so today's talk is going to proceed in three parts.
[04:24.160 --> 04:28.480]  First, laying out the NSA's domestic bulk surveillance authority.
[04:28.480 --> 04:30.780]  I suspect that's going to be recapped for a lot of you.
[04:30.840 --> 04:32.360]  But I want to make sure the playing field's level.
[04:32.880 --> 04:40.900]  Then I'm going to talk about our work on the NSA's international Internet traffic program under Section 702 of the FISA Amendments Act.
[04:40.900 --> 04:46.820]  And then last turn to the phone metadata program under Section 215 of the USA PATRIOT Act.
[04:47.880 --> 04:52.000]  So, I'm going to start with that overview of domestic bulk surveillance programs.
[04:52.160 --> 05:01.440]  And first talk about the earliest court sanctioned program no longer in existence, looking at bulk email metadata.
[05:01.440 --> 05:07.470]  So, the legal constraints on this program were that the NSA was only allowed to acquire metadata.
[05:08.340 --> 05:12.240]  There needed to be reasonable and articulable suspicion to look at the data set.
[05:12.340 --> 05:18.800]  RIS is a legal standard that's used for so-called Terry stops, stopping you on the street.
[05:19.260 --> 05:23.500]  The controversial New York stop and frisk policy was based on RIS.
[05:23.500 --> 05:24.620]  So, it's a pretty low standard.
[05:26.040 --> 05:30.480]  And so, those are the legal constraints for taking a look at email metadata.
[05:31.300 --> 05:37.960]  The source of legal authority there was not the FISA business records provision you might have heard of.
[05:38.040 --> 05:39.560]  Section 215 of the PATRIOT Act.
[05:39.680 --> 05:40.400]  It's been in the news a lot.
[05:40.500 --> 05:42.860]  It was actually Section 214 of the PATRIOT Act.
[05:43.060 --> 05:45.280]  The pen register trap and trace device provision.
[05:45.860 --> 05:55.940]  The reason that's important in NSA overhaul debates is if 214 can get used in the same ways as 215, then 214 also needs some revisions.
[05:56.380 --> 06:02.540]  Not coincidentally, the NSA has taken the position that, well, if there's going to be action on 215, why don't we leave 214 in place?
[06:03.760 --> 06:14.160]  The Fourth Amendment also provides some constraint here since metadata, at least under the law as the NSA views it, isn't protected by the Fourth Amendment.
[06:14.160 --> 06:18.040]  I'm happy to take questions on the specifics of that doctrine in the Q&A, by the way.
[06:18.600 --> 06:20.480]  Okay, so a quick sketch of how this program worked.
[06:20.860 --> 06:41.700]  So, if the laptops represent end user devices and the servers represent email servers, a user might use webmail over HTTP or send a message over SMTP to their email service, and then goes over SMTP to another email service, and then finally someone will retrieve it by webmail or one of these mail retrieval protocols.
[06:42.980 --> 06:57.480]  And while your session with your service might have been encrypted and authenticated, and the recipient's session with their service might have been encrypted and authenticated, for quite some time, the traffic between these servers wasn't encrypted and authenticated,
[06:57.540 --> 06:59.440]  and so the NSA was having a bit of a field day.
[06:59.840 --> 07:01.860]  Thankfully, this is now changing.
[07:03.260 --> 07:06.740]  Large email service providers are now starting to encrypt and authenticate their traffic.
[07:06.740 --> 07:13.160]  So it's not clear this program could even be as effective now as it was historically at collecting information.
[07:13.400 --> 07:15.120]  But anyway, that was the design of the program.
[07:16.480 --> 07:22.360]  And so just a quick sketch of what this looked like in SMTP to give you a sense of how easy it is to pull this data out.
[07:22.620 --> 07:26.520]  If this is an SMTP session, the NSA would get the stuff in red.
[07:27.020 --> 07:29.700]  Pretty easy to just parse that out of traffic over the wire.
[07:30.100 --> 07:32.280]  So that's the first program.
[07:32.420 --> 07:33.320]  It's no longer in existence.
[07:33.320 --> 07:36.120]  We didn't study it primarily because it's no longer in existence.
[07:39.200 --> 07:41.860]  So I thought at minimum you should be aware of it.
[07:42.360 --> 07:50.440]  The next one that was in the news quite a lot actually very recently is the NSA's one-end foreign Internet traffic program.
[07:50.640 --> 07:52.480]  Appears to have started in about 2001.
[07:53.260 --> 07:55.680]  Transitioned to court oversight in about 2007.
[07:56.280 --> 08:00.480]  And I think the NSA's own slides probably provide the best explanation of it.
[08:03.140 --> 08:07.660]  So lots of Internet traffic flows through the U.S.
[08:08.020 --> 08:13.140]  Even if it's not necessarily directed into the U.S. or out of the U.S.
[08:14.460 --> 08:18.820]  So the NSA is in a good position to observe network flows as they move around the Internet.
[08:19.800 --> 08:23.540]  The program I'm talking about is the one known as upstream, not PRISM.
[08:23.540 --> 08:29.740]  They both arise under the same legal provisions, but they operate very differently.
[08:30.460 --> 08:33.640]  So upstream is Internet traffic as it flows over the wire.
[08:33.780 --> 08:37.080]  PRISM is targeted orders at particular services for particular accounts.
[08:37.400 --> 08:38.620]  So upstream is bulk.
[08:38.740 --> 08:39.420]  PRISM is not.
[08:40.760 --> 08:53.840]  So the upstream program was a very controversial part of the recent Privacy and Civil Liberties Oversight Board report on FAA 702.
[08:54.800 --> 09:01.720]  And if you want to get a little more detail on how it operates, that's probably the best place to look.
[09:01.720 --> 09:11.780]  Unfortunately, the details are still incredibly sketchy, including some specifics like is it all one and four in traffic?
[09:11.840 --> 09:17.920]  Is it buffered for a while so that the NSA can look back for a few days for traffic of interest?
[09:19.900 --> 09:24.880]  Are the rules different for organizations versus human persons when they're targeted?
[09:25.920 --> 09:35.500]  Are the rules different when the traffic is some sort of messaging that happens to be going over some corporate network, but that's sort of an intermediary step?
[09:35.740 --> 09:36.860]  It's not really clear.
[09:37.180 --> 09:40.780]  The FISA court certainly has not parsed this all out and the PCLAB didn't either.
[09:41.940 --> 09:46.820]  Here's a quick sketch of what the sort of information that's available from the system looks like.
[09:47.940 --> 09:54.860]  So at least once it gets to an analyst, it looks just an awful lot like ordinary Internet traffic here.
[09:54.860 --> 09:57.100]  Here's a parsed up HTTP session.
[09:58.440 --> 10:01.500]  So again, the legal constraint is that this has to be one and four in.
[10:01.620 --> 10:04.680]  That is, one end of the communication has to be outside the United States.
[10:04.940 --> 10:09.400]  Not entirely clear what the NSA considers to be sufficient for one and four in.
[10:11.380 --> 10:14.460]  The authority is FAA section 702.
[10:14.620 --> 10:19.200]  So every year, the FISA court stamps a targeting order.
[10:19.200 --> 10:28.580]  Then the NSA acting through the FBI, it appears, serves out orders without court involvement on major ISPs.
[10:28.840 --> 10:30.740]  And they provide access to this traffic.
[10:31.180 --> 10:37.440]  The Fourth Amendment in the NSA's view doesn't apply because of one of two doctrines.
[10:37.800 --> 10:42.520]  One being around non-U.S. persons outside the United States aren't entitled to Fourth Amendment protection.
[10:43.620 --> 10:45.640]  It's not clear how good the precedent is.
[10:45.760 --> 10:47.720]  But at any rate, that's the NSA's view of precedent.
[10:48.200 --> 10:52.000]  And the other being the border searching doctrine under the Fourth Amendment.
[10:52.460 --> 10:56.220]  So that's why the NSA believes this program is lawful.
[10:57.960 --> 11:01.400]  Phone calls have certainly been in the news.
[11:01.400 --> 11:05.340]  One of the earliest of the very controversial NSA programs.
[11:06.640 --> 11:11.820]  The legal constraints here look an awful lot like the legal constraints around the email metadata program.
[11:12.140 --> 11:13.100]  Again, metadata only.
[11:13.260 --> 11:14.720]  Again, RAS for querying.
[11:15.300 --> 11:17.320]  There's contact chaining allowed here.
[11:17.860 --> 11:21.300]  Contact chaining also is allowed for the email program, by the way.
[11:22.140 --> 11:27.200]  The idea being you don't just look at the metadata of a target, but you can look a few degrees removed.
[11:28.060 --> 11:29.640]  In the news, very controversial.
[11:29.640 --> 11:35.040]  And this is the program that's under Section 215 of the PATRIOT Act, the so-called business records provision.
[11:35.220 --> 11:38.060]  So not quite the same legal authority as email metadata.
[11:39.120 --> 11:41.760]  The Fourth Amendment rationale, by the way, the exact same.
[11:41.860 --> 11:42.900]  Metadata not protected.
[11:43.960 --> 11:49.000]  There's a financial records program under Section 215 of the PATRIOT Act that the CIA operates.
[11:49.160 --> 11:51.880]  Details on that are pretty thin, too.
[11:52.520 --> 11:53.960]  So we know it exists.
[11:54.120 --> 11:57.140]  There's not so much more that's been published about it yet.
[11:57.940 --> 12:02.360]  There was a trial of a phone location program under Section 215 of the PATRIOT Act.
[12:03.700 --> 12:04.980]  That was short-lived.
[12:05.820 --> 12:07.260]  Allegedly a little narrow in focus.
[12:08.400 --> 12:12.320]  The agency appears to have not thought it was getting valuable information out of it.
[12:12.880 --> 12:15.860]  But certainly believes it has the legal authority to collect that information.
[12:16.840 --> 12:16.960]  Okay.
[12:17.380 --> 12:17.880]  So that's...
[12:17.880 --> 12:19.400]  Those are the major programs we know about.
[12:19.560 --> 12:20.380]  What don't we know about?
[12:20.500 --> 12:20.780]  Who knows?
[12:21.560 --> 12:24.360]  There certainly could be other 215 programs.
[12:24.520 --> 12:25.820]  There could be other 214 programs.
[12:25.820 --> 12:28.140]  There could be other things 702 is getting used for.
[12:29.280 --> 12:37.380]  There's been still very little transparency on the legal reasoning and application of that reasoning side of the NSA debate.
[12:38.240 --> 12:41.320]  So this was our best attempt to synthesize the playing field.
[12:42.600 --> 12:46.460]  And with that out of the way, let me turn to the research we did.
[12:46.460 --> 12:50.000]  But since, for us anyways, this was the much more exciting part.
[12:52.180 --> 12:56.460]  So, starting with the Internet traffic program.
[12:57.300 --> 12:59.900]  The 702-1-4 Internet traffic program.
[13:01.210 --> 13:06.400]  We started by looking through assumptions we found in documents justifying the program.
[13:06.820 --> 13:08.300]  Documents out of the FISA court.
[13:08.440 --> 13:09.480]  Documents out of politicians.
[13:09.800 --> 13:12.400]  Documents out of senior executive branch officials.
[13:12.400 --> 13:17.160]  And started documenting what the assumptions were surrounding the programs.
[13:17.360 --> 13:19.920]  And one was that Americans don't send data outside the U.S.
[13:20.520 --> 13:23.480]  And so we thought, okay, let's see if we can measure that.
[13:23.800 --> 13:25.340]  That seems like a good technical question.
[13:25.440 --> 13:27.480]  Do you send your information outside the U.S.?
[13:29.680 --> 13:34.680]  Another assumption is that, well, when you send information outside the U.S., you know you're doing it.
[13:34.680 --> 13:36.720]  So it's sort of an assumption of the risk argument.
[13:37.260 --> 13:43.520]  Kind of like when you decide to fly internationally, you know you might be subjected to some sort of heightened search at the border.
[13:44.520 --> 13:46.540]  I'm not saying this is great legal reasoning.
[13:46.720 --> 13:49.180]  But at any rate, it was one of the assumptions.
[13:49.540 --> 13:54.680]  And so we thought, okay, let's see if we can find out whether Americans knowingly send their data internationally.
[13:57.360 --> 14:02.240]  And so we found the answer to both questions was, yeah, Americans send data outside the U.S. all the time.
[14:02.240 --> 14:03.940]  And they don't know they're doing it very often.
[14:04.780 --> 14:06.660]  Seemingly wouldn't know they're doing it very often.
[14:07.260 --> 14:09.220]  So let me give a few examples of how this happens.
[14:09.540 --> 14:23.400]  How stuff that would seem like it should be entirely domestic, should seem like it should be entirely protected by the Fourth Amendment, and should have been unambiguously protected by the Fourth Amendment for decades, may actually come within the NSA's Warrantless Surveillance Authority.
[14:24.200 --> 14:27.860]  So first, domestic websites often include foreign content.
[14:28.180 --> 14:30.240]  So here's a trivial example.
[14:30.600 --> 14:35.140]  Let's suppose you're reading a news website and some advertisement loads from some advertising company.
[14:35.320 --> 14:38.100]  And the news website is located inside the United States.
[14:38.100 --> 14:52.980]  Well, the NSA is going to need to go, well, I should say the FBI is going to need to go, and get a wiretap order to monitor your traffic to that website, and the content of the traffic.
[14:53.380 --> 15:00.020]  But suppose the website is in the U.S., but the advertising company isn't in the U.S.
[15:00.220 --> 15:03.820]  There's a sort of echo that goes to the advertising company outside the U.S.
[15:03.820 --> 15:07.260]  With your HTTP request, there's this field included, or a FUR.
[15:07.580 --> 15:14.760]  You know, of course, that the folks who designed the HTTP spec weren't English majors, because now we're stuck with a misspelled referrer in every HTTP request.
[15:21.820 --> 15:29.040]  So by watching an international advertising company, the NSA could very easily see, okay, you've just read something on this news website.
[15:29.280 --> 15:35.660]  And so long as it's a public URL, it doesn't even matter that the NSA can't see content, because, of course, they could go check out the URL themselves.
[15:36.140 --> 15:45.980]  So here's a really simple way in which, just because of the way the web works, the NSA may have the ability to see what you're reading on popular websites.
[15:46.340 --> 15:52.120]  And let me give you what was unambiguously my favorite example of embedded third-party content on a domestic website.
[15:52.480 --> 15:56.660]  So you'd think, like, the House of Representatives website would be as American as Apple Pie.
[15:56.660 --> 16:03.220]  It turns out that there was a little text widget, like a read-aloud widget, that was hosted off of a server in London.
[16:04.700 --> 16:09.020]  Entirely noble aims, but, you know, for folks who had difficulty reading English.
[16:09.720 --> 16:16.040]  But the net effect was that every time you visited a page on the main House of Representatives website, your traffic flew over to London.
[16:16.620 --> 16:18.220]  One-end foreign, pretty clearly.
[16:19.920 --> 16:20.400]  Okay.
[16:20.400 --> 16:27.700]  So we did a crawl looking at some popular websites to get a sense of how pervasive this sort of international traffic leakage was.
[16:28.740 --> 16:30.220]  And we found it all over the place.
[16:30.340 --> 16:32.320]  There were, you know, across all sorts of categories.
[16:32.520 --> 16:36.760]  Like Nike, for example, had a Japanese analytics company they used.
[16:37.340 --> 16:40.260]  There were ads getting loaded from outside the U.S.
[16:40.260 --> 16:47.740]  Some medical providers hosted or embedded little logos certifying that they complied with various legal regimes.
[16:47.840 --> 16:50.320]  Sometimes those logos were hosted on servers outside the U.S.
[16:50.540 --> 16:54.120]  So pretty eclectic mix of what caused the traffic to leak.
[16:54.280 --> 16:56.840]  But clearly it's the case that the traffic leaks a lot.
[16:58.280 --> 17:04.260]  And so it looked to us like very roughly about 10% of popular web pages include this phenomenon.
[17:04.580 --> 17:08.920]  And so that's kind of the net effect of the NSA's legal theories.
[17:08.920 --> 17:13.700]  That, well, take kind of a 10% haircut on the Fourth Amendment.
[17:14.740 --> 17:18.560]  Because that's the proportion of your web traffic that can be surveilled without a warrant.
[17:20.340 --> 17:20.820]  Okay.
[17:21.300 --> 17:24.560]  So that's one way in which traffic might leave the U.S.
[17:24.760 --> 17:27.560]  Another is domestic organizations might just have foreign hosting.
[17:28.020 --> 17:30.440]  So another fun example we stumbled across.
[17:30.620 --> 17:34.440]  GM, you would imagine, you know, very American, headquartered in Detroit.
[17:34.700 --> 17:37.540]  Heck, owned in large part by the American taxpayers for a while.
[17:38.740 --> 17:41.180]  It turns out a lot of their corporate IT is based out of Frankfurt.
[17:41.640 --> 17:46.500]  And so depending on which GM URL you go to, your traffic might actually get bounced over to Frankfurt.
[17:47.220 --> 17:51.360]  And so the NSA could see, you know, which caddy you were checking out.
[17:52.680 --> 17:55.840]  Let me give another example I stumbled across about a week ago.
[17:56.600 --> 17:58.480]  And that is the...
[17:59.620 --> 18:00.100]  Yeah.
[18:00.620 --> 18:01.120]  So, right.
[18:01.220 --> 18:02.160]  Here we are in New York.
[18:02.280 --> 18:06.240]  HOPE, of course, sponsored by 2600, an American organization.
[18:06.600 --> 18:07.740]  Let's do a quick demo here.
[18:09.080 --> 18:10.940]  So make sure this is working.
[18:16.900 --> 18:17.380]  Excellent.
[18:18.720 --> 18:19.200]  Okay.
[18:19.200 --> 18:19.320]  Okay.
[18:19.840 --> 18:20.820]  Let me...
[18:21.820 --> 18:22.320]  Sorry.
[18:23.080 --> 18:23.580]  Yeah.
[18:23.940 --> 18:24.300]  Okay.
[18:29.640 --> 18:30.140]  Okay.
[18:30.280 --> 18:32.700]  So I'm just going to load up the hope schedule.
[18:32.940 --> 18:33.940]  Send that over here.
[18:35.180 --> 18:35.680]  Okay.
[18:38.600 --> 18:39.200]  All right.
[18:39.340 --> 18:39.740]  So...
[18:42.630 --> 18:44.130]  Hopefully that looks okay.
[18:44.570 --> 18:44.850]  Sorry.
[18:44.950 --> 18:46.030]  I'm going to freeze down for a second.
[18:46.630 --> 18:57.990]  So you can see right here, log.x.hook.net is getting an HTTP request that's pulling in an RSS feed of who's going to talk at hope.
[18:58.410 --> 19:02.890]  And so now, if we come back over here.
[19:26.020 --> 19:26.900]  Working on it.
[19:40.940 --> 19:41.940]  Hopefully that worked.
[19:46.600 --> 19:47.500]  Not so much.
[19:47.680 --> 19:47.820]  Okay.
[19:48.140 --> 19:48.660]  Just a second.
[19:49.240 --> 19:49.720]  Nope.
[19:50.080 --> 19:50.100]  Okay.
[19:50.400 --> 19:51.660]  All right.
[19:56.600 --> 19:57.920]  I'm tethering off my cell phones.
[20:03.880 --> 20:07.820]  So anyways, you'll see that this is not a server actually located in the United States.
[20:08.300 --> 20:12.380]  Kind of didn't really hide the punchline too much on this.
[20:13.740 --> 20:14.780]  Good job.
[20:16.800 --> 20:18.560]  This is like the most anti-climactic.
[20:32.120 --> 20:33.280]  Just more lead.
[20:33.940 --> 20:34.400]  The U.S.
[20:34.920 --> 20:35.380]  Wow.
[20:36.000 --> 20:37.920]  This is why Verizon is so great.
[20:38.180 --> 20:38.360]  Yeah.
[20:41.440 --> 20:41.960]  Okay.
[20:42.140 --> 20:42.660]  Very excited.
[20:42.840 --> 20:43.420]  Love the U.S.
[20:48.260 --> 20:50.100]  Houston, we have warrantless lift off.
[20:51.500 --> 20:52.060]  Okay.
[20:55.600 --> 20:56.780]  So you get the idea.
[21:02.820 --> 21:05.020]  So things that would seem very, very domestic.
[21:05.380 --> 21:06.840]  Not necessarily so domestic.
[21:08.500 --> 21:12.100]  Also, there can be services that are not obviously foreign.
[21:12.320 --> 21:14.840]  So you might think you're using something that's based in the U.S.
[21:14.880 --> 21:15.220]  It's a U.S. company.
[21:15.780 --> 21:16.440]  Very popular.
[21:16.620 --> 21:17.740]  Just happens not to be.
[21:18.240 --> 21:21.780]  So, for example, if you've ever scheduled a conference with the website Doodle.
[21:22.400 --> 21:23.860]  Turns out they're a Swiss company.
[21:24.300 --> 21:25.940]  And a lot of their infrastructure is out of Switzerland.
[21:26.400 --> 21:32.380]  So your conferences getting scheduled through Doodle are apparently open to warrantless wiretapping.
[21:34.580 --> 21:36.600]  Spotify, popular music service.
[21:37.060 --> 21:38.320]  Also not in the U.S.
[21:38.640 --> 21:41.900]  Is.gd, popular URL shortener.
[21:42.000 --> 21:42.780]  You might use it on Twitter.
[21:43.560 --> 21:48.280]  Turns out every time someone goes through one of those is.gd links, they're hitting up a server in the U.K.
[21:48.280 --> 21:55.500]  And actually, some people who... some folks use is.gd links in a sort of ironic way.
[21:55.800 --> 21:59.000]  So I don't know if this is like actually Bruce Schneier's official feed.
[22:00.900 --> 22:08.400]  But anyways, kind of funny that like you click a link to learn about the NSA and there goes your traffic into the one end foreign bucket.
[22:11.120 --> 22:11.680]  Okay.
[22:11.680 --> 22:18.840]  So fourth way in which your traffic might get... wind up outside the U.S. in a way you might not expect.
[22:20.240 --> 22:22.220]  Sometimes routes just go all over the place.
[22:22.600 --> 22:26.740]  So for example, here's a route that occurred not too long ago.
[22:27.480 --> 22:31.840]  That's from within the same city to within the same... two points within the same city.
[22:32.600 --> 22:34.600]  So the Internet can be a little unpredictable.
[22:35.080 --> 22:38.220]  And so just by virtue of how routing works, your traffic might leak.
[22:38.920 --> 22:44.880]  So the punchline to all of this is that ordinary Americans are known to send tons of their traffic outside the U.S.
[22:45.060 --> 22:55.040]  Those assumptions underlying the legal arguments for this program are, I think, all the more questionable in light of...
[22:55.980 --> 22:58.040]  In light of what we found.
[22:58.440 --> 23:04.660]  And so I think it suggests that just this one end foreign rule is not really consistent with how information technology works today.
[23:05.120 --> 23:14.460]  And actually, this particular study got cited by the DNI review group for the idea that maybe one end foreign doesn't make so much sense anymore.
[23:15.320 --> 23:16.660]  I think I saw a hand up.
[23:16.760 --> 23:19.680]  I'm planning to take questions at the end, planning to leave some time for them.
[23:20.020 --> 23:22.120]  So please just hang on to them.
[23:23.040 --> 23:23.500]  Okay.
[23:24.100 --> 23:26.520]  So now let me turn to our phone metadata study.
[23:27.520 --> 23:30.940]  And again, start by unpacking the assumptions baked into it.
[23:31.920 --> 23:39.780]  One is that metadata isn't identifiable, which I think was kind of laughable to a lot of the folks we talked to about it.
[23:39.900 --> 23:42.100]  We thought, okay, like, we'll see if we can identify phone metadata.
[23:42.280 --> 23:43.600]  We're pretty confident we could do that.
[23:44.840 --> 24:01.780]  There's also this assumption that metadata wouldn't be densely interconnected, that by imposing limits on how long the data was retained, by imposing limits on how many hops you could make throughout the call graph, that would prevent getting too many folks' metadata swept up in queries.
[24:02.320 --> 24:05.600]  And so we thought, okay, let's see if we can figure out what the structure of the phone call graph looks like.
[24:07.920 --> 24:10.780]  Another assumption baked into this program is that, well, it's just metadata.
[24:10.980 --> 24:12.140]  It can't be that sensitive.
[24:12.280 --> 24:12.960]  What's the big deal?
[24:13.960 --> 24:15.860]  We are pretty skeptical of this one, too.
[24:16.920 --> 24:21.180]  And so we thought, all right, let's see if we can draw some sensitive inferences out of phone metadata.
[24:22.260 --> 24:33.380]  And so the starting problem we had to overcome in our study was how on earth were we going to actually get our hands on some phone metadata, and some ground truth to see if the inferences we were drawing were accurate.
[24:34.340 --> 24:37.080]  And so our first thought was, like, maybe let's ask some telecoms.
[24:37.620 --> 24:39.860]  And yeah, that's like pretty much exactly what happened.
[24:40.500 --> 24:42.360]  Yeah, they were not so forthcoming.
[24:43.700 --> 24:48.080]  And so we thought, okay, like, why don't we ask the Internet?
[24:48.080 --> 24:50.640]  Like, pretty please, Internet, we're trying to study the NSA.
[24:50.640 --> 24:53.820]  Is there any chance you would trust us with your phone metadata?
[24:55.540 --> 24:56.880]  And a bunch of folks did.
[24:57.560 --> 25:02.020]  So we put this app up on Google Play called Metaphone.
[25:02.400 --> 25:05.080]  And here's a quick sketch of what the screens looked like.
[25:05.580 --> 25:08.040]  So, you know, you bet we had a disclosure.
[25:08.480 --> 25:10.560]  Went back and forth with IRB on this one.
[25:13.100 --> 25:15.840]  And then folks could connect their Facebook account.
[25:17.560 --> 25:19.760]  Right, so this was how we solved the ground truth problem.
[25:19.920 --> 25:26.420]  That we said, you know, give us access to your Facebook account, and that way you don't have to answer a bunch of survey questions about your age, your gender, your location.
[25:26.920 --> 25:30.320]  Facebook's API will provide that to us in a well-structured way.
[25:34.040 --> 25:37.760]  This was all directed on the U.S. Google Play store.
[25:37.980 --> 25:43.500]  And actually, I think we occasionally got negative reviews on Google Play saying, I want to give you my information.
[25:44.000 --> 25:46.020]  I'm outside the U.S., why won't you let me?
[25:48.200 --> 25:48.660]  Yeah.
[25:50.040 --> 25:51.700]  So a lot of folks participated.
[25:52.260 --> 25:58.140]  We had the good fortune to get some press attention on this as we shared some parts of our results.
[25:58.140 --> 26:03.320]  So each of the red lines represents some amount of media attention around some release.
[26:03.520 --> 26:06.480]  And then the blue bars are how many folks submitted data on each week.
[26:06.600 --> 26:10.740]  So kind of got a little spike as we shared our results.
[26:13.220 --> 26:16.560]  The sample is definitely not representative of the U.S. population.
[26:16.780 --> 26:17.920]  You have an Android phone.
[26:18.020 --> 26:19.140]  You have to have a Facebook account.
[26:19.380 --> 26:20.920]  You had to opt into this.
[26:21.980 --> 26:24.980]  I'm pretty confident some of these people were lying about their age.
[26:27.920 --> 26:34.200]  And then, of course, the gender imbalance was, unfortunately, well, roughly like, you know, DEFCON.
[26:36.760 --> 26:40.920]  I was going to say hope, but I've been really heartened to see that hope is not DEFCON.
[26:42.360 --> 26:50.080]  So we had all told about 825 participants, which leads to some awkward conversations with folks in Silicon Valley.
[26:50.300 --> 26:51.340]  Like, did your app go big?
[26:51.440 --> 26:54.620]  And we're like, yeah, we were really impressed with how many folks were willing to give us their metadata.
[26:54.940 --> 26:56.320]  They were like, you know, hundreds of thousands?
[26:56.540 --> 26:57.140]  Not quite.
[26:59.300 --> 27:03.980]  But we thought more than enough to draw some conclusions about the privacy properties of metadata.
[27:04.300 --> 27:08.460]  About 250,000 calls, one and a quarter million text messages in our data set.
[27:08.800 --> 27:13.040]  So a good amount for making some observations.
[27:14.920 --> 27:19.780]  So we looked into this idea that phone metadata isn't identifiable.
[27:20.600 --> 27:26.500]  It's an idea that's actually gotten tossed out by some folks who really should have vetted it more thoroughly.
[27:26.720 --> 27:28.960]  So for example, here's what President Obama had to say.
[27:29.720 --> 27:32.700]  Here is what Senator Feinstein had to say.
[27:33.660 --> 27:46.860]  And on the other side, Judge Leon from the District of Columbia, the most confusingly titled court, there are two districts, said, gee, it sure seems like it would be easy to identify these.
[27:46.860 --> 27:53.620]  And then Ed Felton at Princeton did a declaration for the ACLU pointing up it sure seems like it would be easy to identify this stuff.
[27:54.080 --> 28:03.520]  So not to like, you know, kind of ruin the punchline, but shock, the computer scientists were right and the politicians were wrong.
[28:04.480 --> 28:10.080]  So, okay, we sampled 5,000 numbers and hit up a bunch of APIs.
[28:11.180 --> 28:18.380]  Here are three of the ones we played with, trying to see if we could match those numbers to identities of individuals and businesses.
[28:19.500 --> 28:22.300]  And so here are the hit rates we got on each of these.
[28:22.680 --> 28:28.080]  All told, we got a little over 27% with just free public APIs.
[28:28.720 --> 28:35.560]  And then to see what a little elbow grease could get us, we just sampled 100 numbers randomly out of our data set.
[28:35.560 --> 28:40.920]  And then just using Google, just like popping the number into Google.
[28:41.260 --> 28:48.180]  And a free commercial database, we got to 91 of them identified in addition to the sources with APIs.
[28:48.560 --> 28:54.060]  So, the notion that phone numbers aren't identifiable is definitely not true.
[28:55.040 --> 28:55.600]  Okay.
[28:56.960 --> 29:07.620]  So, we next looked at the structure of the call graph, trying to understand what the NSA's durational and hop limitations did in effect.
[29:08.140 --> 29:20.080]  And to understand how those limitations work, it helps to give a quick sketch of how the 215 program works as a sort of database, as a sort of shuffling data between databases project.
[29:20.080 --> 29:25.040]  So, there's this raw phone metadata that comes in from the phone companies.
[29:25.700 --> 29:27.980]  And individual analysts can't look at it.
[29:28.460 --> 29:35.840]  Technicians can, for purposes of combing over it, and making sure it's well formatted, making sure the agency's not getting fields it's not supposed to get.
[29:36.280 --> 29:39.100]  But individual line-level analysts aren't supposed to look at it.
[29:40.040 --> 29:47.240]  Separately, there is a list of seed queries that have been cleared for reasonable and articulable suspicion.
[29:47.440 --> 29:48.960]  Originally, those were cleared internally.
[29:48.960 --> 29:51.060]  Now, they're all cleared with the FISA court.
[29:53.120 --> 29:56.260]  And then, those two can be combined.
[29:56.640 --> 30:08.140]  A seed query in the raw data, plus a couple of degrees, hops, outwards in the raw data, to get a sort of neighborhood around some seed phone number.
[30:08.400 --> 30:12.780]  And then, finally, the results get put in some database called the corporate store.
[30:12.960 --> 30:14.640]  Why it's called the corporate store is beyond me.
[30:15.280 --> 30:27.900]  Once data is in the corporate store, it can then be analyzed without any further need for looking for RIS, and with some more substantially laxer constraints than the raw data has.
[30:28.740 --> 30:36.500]  So, there's this incentive to put as much data as possible in the corporate store if what you'd like to do is maximize the information available to line-level analysts.
[30:38.160 --> 30:46.900]  So, in thinking through how much information might be available for a given seed query, we started by thinking, okay, so there's two or three hops out.
[30:47.620 --> 30:50.900]  Three initially, now the NSA has voluntarily scaled it back to two.
[30:52.460 --> 30:57.220]  And so, here's just kind of a recap of what the hops look like.
[30:57.420 --> 31:17.940]  And for clarification, as best I can tell, at the last hop, the agency collects the metadata of each number that's within that hop count, such that you're actually kind of three plus or two plus hops away, because you can see numbers that are connected to the second hop,
[31:18.020 --> 31:19.240]  or that are connected to the third hop.
[31:20.860 --> 31:23.880]  One call in the past five years is currently enough for a hop.
[31:24.420 --> 31:31.760]  Under the White House's reform proposal, that would get scaled back to about 18 months, which is the retention requirement for phone companies.
[31:32.580 --> 31:33.020]  Anyways.
[31:34.460 --> 31:34.900]  Okay.
[31:35.180 --> 31:38.380]  So, we thought we should ask, how much of the call graph could go into the corporate store?
[31:38.580 --> 31:43.540]  And our first take on this was, well, let's just kind of think about it like a Kevin Bacon problem.
[31:44.060 --> 31:57.260]  And a pretty diffuse social network, where very roughly speaking, you have some degree associated with each node, and you multiply out the degree, and there you have roughly how many folks are connected within some number of hops.
[31:57.640 --> 31:59.920]  And we certainly weren't the only ones with that intuition.
[32:00.240 --> 32:07.000]  And you would get some number of folks from the thousands, the millions, depending on what you assume the degree was, and how many hops you went out.
[32:08.240 --> 32:12.600]  So, we thought, okay, that's kind of what this is going to look like.
[32:12.820 --> 32:16.120]  We're going to be able to say something about the rough degree of a phone subscriber.
[32:16.800 --> 32:20.640]  We're not going to be able to say much more of that, because our call graph is going to be really disconnected.
[32:21.060 --> 32:26.860]  And we thought it was going to be really disconnected, because, well, we have a scattershot sample of folks who just decided to give us their phone metadata.
[32:27.440 --> 32:28.680]  It's a time-limited sample.
[32:28.680 --> 32:33.980]  We only got about 180 days from most participants, as opposed to the NSA's five years.
[32:34.180 --> 32:43.460]  And it was, you know, all things considered a small sample, enough that we thought we could draw inferences, but, you know, it was hardly a, you know, hardly Silicon Valley big.
[32:45.080 --> 32:48.460]  So that's what we thought we were going to get, and it turned out we were entirely wrong.
[32:49.320 --> 32:55.180]  And so let me give you some diagrams of our connectivity graph.
[32:55.420 --> 32:58.960]  The blue dots represent participants in our study.
[32:59.180 --> 33:03.480]  The gray lines represent some connectivity between the dots.
[33:04.820 --> 33:09.200]  And then the red dots are going to represent non-participant numbers.
[33:09.200 --> 33:12.880]  I'm going to add those in with the highest degree ones first.
[33:13.860 --> 33:20.040]  So you can see here there are just two people in this visualization who had directly connected with each other.
[33:20.180 --> 33:23.800]  So not a closely connected participant base.
[33:23.800 --> 33:27.740]  So the highest degree hub we found was T-Mobile's voicemail.
[33:27.920 --> 33:35.300]  Turns out if you have a phone configured to dial into voicemail, well, there's this, there's a hub.
[33:35.420 --> 33:40.960]  So everyone who has T-Mobile voicemail is just two hops away from everyone else who has T-Mobile voicemail.
[33:42.380 --> 33:44.640]  Yeah, so that's quite a lot of connectivity.
[33:45.280 --> 33:47.320]  That doesn't look a lot like a diffused social network.
[33:47.400 --> 33:51.940]  That looks a lot like a hub-and-spoke network where just a couple of hops can actually get you pretty far.
[33:52.620 --> 33:56.280]  And, you know, adding some more in, customer service.
[33:56.420 --> 33:58.980]  I guess a lot of folks have a need to call Comcast customer service.
[34:01.660 --> 34:08.020]  Skype uses certain predictable numbers for their outbound calling if you haven't paid for a custom Skype number.
[34:08.440 --> 34:18.280]  So if you've taken a call from someone who uses Skype, then you're connected to everyone else who's taken a call who uses Skype from that auto set Skype number.
[34:19.040 --> 34:20.960]  Lots of calls with 611.
[34:21.720 --> 34:23.320]  Let me add some more in here.
[34:23.920 --> 34:27.700]  So then things started to get really fun when we started to see the telemarketers in our data set.
[34:28.600 --> 34:32.040]  So they're like just making the NSA's job a lot easier, right?
[34:32.140 --> 34:37.660]  They're just trying to dial a bunch of phone numbers which has the net effect of putting them all two hops from each other.
[34:38.120 --> 34:38.520]  Yeah.
[34:41.020 --> 34:41.420]  Right.
[34:42.160 --> 34:50.060]  So you see how more and more of these hub numbers crop up, pretty densely connecting our participant population.
[34:50.400 --> 34:55.040]  And so by the time we're done, it doesn't look like a diffuse social graph at all.
[34:55.240 --> 35:00.180]  Actually, it looks like you might get an awful lot with just a couple of hops.
[35:01.660 --> 35:11.120]  And so to show this slightly differently in graphical form, the horizontal axis here is the hub numbers just ranked.
[35:11.340 --> 35:15.820]  And then the vertical axis is the proportion of participants that are connected.
[35:15.820 --> 35:35.280]  And so one key takeaway from this is instead of thinking about this as a Kevin Bacon problem, where you're kind of multiplying out from a certain number of people who might be connected, maybe another way to think about it is just meaningful share of the entire U.S. population that's within two or three hops of a number.
[35:41.060 --> 35:50.160]  So, right, very easy for a hub to be between some seed and some individual phone user.
[35:50.900 --> 35:57.580]  We found that over 90% of the participants in our study were within a connected component, over 60% within four hops.
[35:58.860 --> 36:08.100]  Thinking through the possible three hop connections, we didn't have a good way of seeing three hop connections because our participants weren't connecting directly to each other.
[36:08.100 --> 36:13.260]  But you could imagine something like one of these two arrangements, right?
[36:13.380 --> 36:21.240]  So you call someone who in turn calls a hub, you're then three hops away from everyone else connected to that hub, or the hub could come earlier in the chain.
[36:25.560 --> 36:41.920]  And so we did a resampling approach to try to roughly get a handle on what the implications of these hubs were for three hop call connectivity, and we found over 50% of our participants were connected within three hops in expectation.
[36:42.560 --> 36:43.640]  So three hops gets an awful lot.
[36:45.360 --> 36:54.460]  And to leave you with one other way in which three hops can be substantially more concerning than two hops, you can even have hubs connecting to hubs, right?
[36:54.460 --> 37:01.020]  So if someone from Skype has at any point called Comcast, you have the Skype hub now connected to the Comcast hub.
[37:01.700 --> 37:05.060]  And so the numbers start to get pretty darn big.
[37:06.340 --> 37:10.720]  So the takeaway for us was that this hop limitation isn't much of a limit.
[37:10.820 --> 37:15.080]  You would have to do some mandatory removal of high degree nodes in the call graph.
[37:15.560 --> 37:26.180]  And even then, the degree of participants, the number of folks they were calling was so high that there was still a large number of people within the neighborhood of two or three hops.
[37:26.880 --> 37:33.780]  So the hop limit didn't seem to us like a particularly persuasive means of cabining metadata surveillance just on the numbers.
[37:34.520 --> 37:39.160]  And one side note to all of this, it's not really clear what moving the raw data solves.
[37:39.160 --> 37:56.180]  If that raw data on the database diagram I showed resided with the phone companies as opposed to residing with the NSA, when so much of that data can get shoved into the NSA's corporate store, it's not really clear what the big privacy win would be by moving all of that data over to phone companies.
[37:57.380 --> 37:57.860]  Okay.
[37:58.140 --> 38:01.280]  So that's all I wanted to say about the graph of the phone network.
[38:02.520 --> 38:07.780]  Now I want to turn to inferring some private information about phone users.
[38:09.480 --> 38:14.580]  And one experiment we ran was trying to identify relationships automatically.
[38:14.940 --> 38:20.920]  And so we took as a fun little weekend task inferring folks' relationship status just from their phone metadata.
[38:21.220 --> 38:27.120]  It turns out if you know someone's in a relationship, figuring out who they're in a relationship with is really trivial.
[38:27.280 --> 38:28.500]  It's like the person they call all the time.
[38:28.900 --> 38:31.540]  But figuring out whether they're in a relationship is trickier.
[38:31.760 --> 38:36.340]  And so we thought, okay, let's see if we can figure it out just using automated graph analysis.
[38:36.340 --> 38:40.780]  And most of these, by the way, were not public relationship statuses.
[38:40.900 --> 38:50.000]  They were shared only with friends or one of the other Facebook settings you can twiddle.
[38:51.980 --> 38:56.660]  So here's the performance we were able to get just with like a weekend of hacking away on this.
[38:56.840 --> 38:59.400]  Not great, but we thought pretty good.
[38:59.860 --> 39:03.940]  And anyways, right, like what do a couple computer scientists do on a weekend instead of like going out on dates?
[39:03.940 --> 39:05.740]  They figure out if other people are going out on dates.
[39:07.240 --> 39:08.200]  So, right.
[39:08.580 --> 39:09.820]  So we got decent performance.
[39:09.980 --> 39:13.060]  We're still hacking away to see if we can do substantially better.
[39:13.220 --> 39:17.380]  But clearly there's a lot of opportunity here for doing automated analysis.
[39:19.340 --> 39:23.280]  We also looked into whether we could draw some pretty sensitive inferences out of phone metadata.
[39:24.280 --> 39:36.060]  And so we went looking for businesses that were associated with sensitive traits and found, for example, there was a participant who had had calls to and from substance abuse treatment center.
[39:37.240 --> 39:42.320]  There was a participant who had calls to and from an STI outpatient center.
[39:43.400 --> 39:50.840]  There was one participant that got a lengthy call from a major medical center, then from a clinical psychologist at that medical center.
[39:51.260 --> 39:56.460]  Then some calls from local pharmacies, calls to and from cardiology groups.
[39:57.000 --> 40:01.780]  And finally regular calls to a home heart reporting line used for a medical device to monitor an arrhythmia.
[40:03.200 --> 40:05.120]  The individual had an arrhythmia.
[40:05.680 --> 40:08.940]  We were actually able to confirm that using public information.
[40:08.940 --> 40:12.780]  So there's finding a medical condition just using someone's phone metadata.
[40:13.060 --> 40:15.400]  Really not challenging.
[40:16.800 --> 40:24.340]  Or there was another participant in our study who made a number of calls to a local handgun sales and repair.
[40:24.560 --> 40:29.240]  I should say handgun and also specialized in a specific kind of assault rifle, the AR-15.
[40:30.240 --> 40:36.720]  Followed by some lengthy calls to a company that specializes in making an AR-15 platform rifle.
[40:36.720 --> 40:42.540]  We were able to confirm again using other information that this individual owns an AR-15 rifle.
[40:42.800 --> 40:48.060]  So they're spotting a specific model of gun that someone owns just using their phone metadata.
[40:49.760 --> 40:57.500]  So we have this hypothesis that maybe in fact phone use is more common in sensitive communications.
[40:57.500 --> 41:08.920]  There are these sectors, whether it's medical practice or legal practice or gun stores, where just as a matter of path dependency, they tend to use the phone a little more than others.
[41:09.280 --> 41:22.420]  And so in fact, maybe this focus on phone metadata can be really quite revealing, perhaps more revealing than other forms of metadata you might get because of its tendency to get used in these sensitive contexts.
[41:23.420 --> 41:27.780]  So that sums up everything I wanted to say about the NSA's programs and about our research.
[41:28.080 --> 41:30.780]  I wanted to close with a quick request to you.
[41:31.100 --> 41:36.600]  And that is that unambiguously the government needs help from you, from our community.
[41:37.420 --> 41:45.340]  Our lab got tons and tons of questions about our work from folks at pretty senior levels in the federal government.
[41:45.340 --> 41:48.460]  Because they didn't have another place to get the facts.
[41:49.040 --> 41:53.940]  And so my request is maybe if we could change the diagram a little bit.
[41:54.100 --> 42:05.380]  And maybe if our community could step in and provide the facts to NSA oversight authorities that the NSA isn't providing right now.
[42:05.580 --> 42:08.760]  That's not to say it's not mutually exclusive with saying screw the NSA.
[42:10.460 --> 42:16.200]  But it's a way of, I think, constructively affecting the debate using the computer science skills we have.
[42:16.780 --> 42:18.880]  And so I guess I would...
[42:19.400 --> 42:24.060]  To the extent it seems of interest to you, I think it would be an extraordinarily productive use of your time.
[42:24.740 --> 42:25.000]  Okay.
[42:25.560 --> 42:27.920]  That is the sum total of my material.
[42:28.120 --> 42:31.100]  And I'm happy to answer Q&A, including beyond the material itself.
[42:31.420 --> 42:31.620]  Thanks.
[42:41.220 --> 43:01.260]  In looking at this surveillance data, were you able also to look internally or analyze the slogan documents in some way, help us understand why this massive investment in surveillance technologies like this happens at this time?
[43:01.420 --> 43:04.060]  Is it just about terrorism or something more?
[43:05.520 --> 43:24.320]  Well, with respect to the legal arguments around these surveillance programs and why they're getting made in ways that seem very uninformed, and why they're getting made that way now, a cynical take would be it's just strategic.
[43:24.580 --> 43:24.940]  Right?
[43:25.060 --> 43:30.000]  If kind of the technical details were known, maybe the FISA court would come down differently.
[43:30.000 --> 43:31.980]  Maybe the White House would feel differently.
[43:33.940 --> 43:35.600]  Maybe DOJ would feel differently.
[43:36.940 --> 43:44.800]  I guess there's a sort of other way of reading it, a sort of banality of legal incompetence version of just...
[43:44.800 --> 43:47.920]  These are a bunch of lawyers who don't have great technical backgrounds.
[43:48.600 --> 43:51.100]  And I'm sure many of you have seen it in practice.
[43:51.120 --> 43:59.720]  If someone doesn't know the right questions to ask of a technical expert, they often don't get details they really wish they would have known.
[43:59.720 --> 44:05.260]  And so I think that is also a version of events that's entirely consistent with these legal rules.
[44:05.360 --> 44:07.040]  Can I just follow up for a second?
[44:07.500 --> 44:15.720]  Because really what I'm getting at is the political rationale for all this as opposed to the legal justification for it.
[44:15.720 --> 44:19.040]  You made reference, for example, to financial records.
[44:19.400 --> 44:22.200]  But there doesn't seem to be a lot of information about this.
[44:22.420 --> 44:25.720]  And that seems to be under the CIA, not the NSA.
[44:26.200 --> 44:31.700]  How do you understand what actually is being monitored here?
[44:31.900 --> 44:39.920]  I don't think I have any special insight into the political dynamics of why the NSA is surveilling in the ways it is.
[44:39.920 --> 44:41.100]  I'm happy to talk about it offline.
[44:41.520 --> 44:43.660]  It seems like there were a bunch of other questions.
[44:43.840 --> 44:44.340]  Let's go...
[44:44.340 --> 44:46.060]  I'm going to just go across like this.
[44:46.440 --> 44:46.560]  Yeah.
[44:46.860 --> 44:55.820]  I just wanted to know what you thought about inferring data via Skype or phone, which is safe.
[44:58.320 --> 45:05.560]  Well, it seems like Skype information is not at present getting passed to the government in bulk.
[45:05.560 --> 45:11.260]  So, at least with respect to metadata, a more targeted order would be required.
[45:14.500 --> 45:15.020]  Yeah.
[45:15.260 --> 45:18.060]  There's certainly the ability to do lawful intercept on Skype now.
[45:18.440 --> 45:20.480]  But the same...
[45:20.480 --> 45:25.280]  So, with respect to a wiretap, it's very similar to a phone call.
[45:25.460 --> 45:26.780]  Go get a wiretap order.
[45:27.520 --> 45:34.300]  If you're curious on the details of this stuff, by the way, the sort of core orders you would need for a particular service to intercept.
[45:35.200 --> 45:42.040]  Stanford Law is going to be launching a class, one of these free online classes, on surveillance law in the fall.
[45:42.400 --> 45:44.940]  So, there's going to be a lot more detail on that there.
[45:46.040 --> 45:46.640]  Yeah?
[45:46.720 --> 45:48.560]  I just wanted to know, do
[45:51.940 --> 45:58.200]  they focus more on computer networks or do they focus more on phone?
[46:00.760 --> 46:05.620]  Well, yeah, I don't have any special insight into the NSA's internal priorities.
[46:06.420 --> 46:18.560]  But with respect to ordinary law enforcement, there's no doubt there's a path of least resistance current and focuses on various surveillance systems.
[46:18.560 --> 46:32.720]  And so, you know, to the extent you're getting investigated by the local PD, you're probably more likely to get surveilled by some particular form of intrusive surveillance, if it's something they've used before they know the process for, and so on.
[46:33.720 --> 46:37.360]  Yeah, my understanding is Skype has gotten fairly routinized at this point.
[46:39.420 --> 46:40.020]  Yeah?
[46:40.680 --> 46:41.040]  Yes?
[46:41.300 --> 46:46.500]  So, it seems to me that something we might want to be concerned about is connection.
[46:47.080 --> 46:55.400]  So, that is, let's suppose we've got someone who, you know, NSA has RIS, the local problem.
[47:08.540 --> 47:12.700]  So, you mean intentionally dialing the RIS number?
[47:15.220 --> 47:24.440]  So, I'm not aware of any, at least, government internal analysis that's been released on what the legal implications of that would be.
[47:24.440 --> 47:24.460]  Okay.
[47:25.740 --> 47:34.380]  For all of the criticism it gets, and all the criticism it deserves, I can't help but imagine the FISA court would not have a sense of humor about that.
[47:35.160 --> 47:43.760]  That seems like pretty overstepping when, you know, you're intentionally circumventing a surveillance limitation solely by technical shenanigans.
[47:44.480 --> 47:45.400]  Yeah, I...
[47:45.400 --> 47:49.660]  But there's no specific crime security in a place to address that?
[47:50.740 --> 47:51.300]  Um...
[47:51.300 --> 47:51.660]  I...
[47:52.360 --> 47:52.540]  I...
[47:52.540 --> 47:53.140]  I...
[47:56.070 --> 47:57.810]  I think that would...
[47:58.370 --> 48:00.490]  I mean, I'm not the NSA's lawyer.
[48:00.750 --> 48:01.470]  Certainly not.
[48:01.910 --> 48:09.670]  If I were working inside the NSA as counsel, I would certainly not, you know, look kindly on an analyst who wanted to do that.
[48:09.670 --> 48:10.230]  Um...
[48:10.230 --> 48:16.450]  I think that would be a very, very strained reading of the FISA court orders that set up the phone metadata surveillance program.
[48:17.170 --> 48:17.750]  Yeah.
[48:46.830 --> 48:52.110]  So, let me start with removal of the hubs and then what the impact of removal of the hubs would be.
[48:52.710 --> 48:53.190]  Um...
[48:53.190 --> 48:53.890]  So, uh...
[48:53.890 --> 48:57.070]  The FISA court has a really weird kind of backwards view on this.
[48:57.330 --> 48:59.130]  Where there is, uh...
[48:59.130 --> 49:01.830]  There is some language in their opinions where they...
[49:01.830 --> 49:03.930]  They view it as a concession to the NSA.
[49:04.370 --> 49:07.610]  That they would allow the NSA to try to identify hub numbers.
[49:07.890 --> 49:09.330]  They don't say the NSA has to.
[49:10.070 --> 49:10.450]  Um...
[49:10.450 --> 49:11.190]  And, uh...
[49:11.810 --> 49:12.250]  And, uh...
[49:12.250 --> 49:19.490]  And their reasoning is, well, we're allowing them to look over the entire data set and generate some analytics in a way that we weren't going to otherwise.
[49:19.490 --> 49:20.310]  Um...
[49:20.310 --> 49:20.450]  Right.
[49:20.530 --> 49:21.690]  Of course, that's like kind of backwards.
[49:21.930 --> 49:24.070]  The whole point of removing the hub numbers is to protect privacy.
[49:24.730 --> 49:24.990]  Uh...
[49:24.990 --> 49:27.470]  So, once again, practicing computer science without a license.
[49:27.790 --> 49:28.010]  Uh...
[49:28.010 --> 49:29.350]  And from the federal bench, no less.
[49:30.270 --> 49:30.710]  Uh...
[49:30.710 --> 49:31.190]  Uh...
[49:31.190 --> 49:31.490]  Uh...
[49:31.490 --> 49:33.050]  So, um...
[49:33.050 --> 49:36.630]  So, the NSA appears to have authority to remove these numbers.
[49:36.830 --> 49:38.970]  As for whether they do, uh...
[49:38.970 --> 49:40.570]  Not entirely clear.
[49:40.750 --> 49:41.850]  And critically, uh...
[49:41.850 --> 49:45.430]  To the extent they do, not clear at which stage of the pipeline they do it.
[49:45.810 --> 49:46.110]  Um...
[49:46.110 --> 49:47.070]  Because, uh...
[49:47.070 --> 49:52.450]  Again, if the incentive is to maximize that corporate store, that database of phone metadata that's not subject to legal...
[49:52.450 --> 49:52.870]  Uh...
[49:52.870 --> 49:53.170]  Substantial...
[49:53.170 --> 49:54.670]  Such substantial legal constraints.
[49:55.310 --> 49:55.770]  Um...
[49:55.770 --> 49:57.090]  Why not detect the hubs?
[49:57.510 --> 49:57.850]  Uh...
[49:57.850 --> 50:01.150]  Then use the hubs for hops in moving everything to the corporate store.
[50:01.370 --> 50:04.210]  Then remove the hubs out of analysis within the corporate store.
[50:04.850 --> 50:05.070]  Um...
[50:05.070 --> 50:05.950]  And so...
[50:05.950 --> 50:06.470]  That...
[50:06.470 --> 50:06.950]  That's...
[50:07.430 --> 50:07.810]  Um...
[50:07.810 --> 50:10.970]  So, in short, the kind of legal structure for removing, uh...
[50:10.970 --> 50:11.170]  The...
[50:11.170 --> 50:12.890]  The hubs is, um...
[50:13.830 --> 50:14.210]  Uh...
[50:14.210 --> 50:14.770]  Unclear.
[50:15.250 --> 50:15.610]  And shaky.
[50:15.770 --> 50:18.670]  As for what the impact would be, it's hard to filter out, uh...
[50:18.670 --> 50:19.770]  Just the hubs, since there...
[50:19.770 --> 50:22.450]  There's this kind of long tail of, like, the local pizza shop.
[50:22.610 --> 50:24.610]  That would be hard to knock out, uh...
[50:24.610 --> 50:25.050]  In a...
[50:25.050 --> 50:28.510]  In a very reliable way without risking losing some intelligence information.
[50:28.810 --> 50:34.350]  So, you might be able to get rid of the really high degree ones, but there'd still be a pretty substantial impact, it seems, on the topography of the graph.
[50:34.350 --> 50:40.110]  And then, even if you were to somehow just get rid of all the hubs and rely on just the diffuse call graph, uh...
[50:40.110 --> 50:41.050]  There's still, uh...
[50:41.050 --> 50:41.890]  A sufficiently high degree.
[50:42.050 --> 50:43.270]  There'd be a lot of information.
[50:43.910 --> 50:44.090]  Um...
[50:44.090 --> 50:44.270]  Okay.
[50:44.370 --> 50:44.950]  So, here's my plan.
[50:45.150 --> 50:45.450]  Uh...
[50:45.450 --> 50:45.950]  We do...
[50:45.950 --> 50:50.230]  Take the remaining questions together, and then I'll answer both, and then we'll call it a day.
[50:50.670 --> 50:50.830]  Uh...
[50:50.830 --> 50:51.390]  So, did...
[50:51.390 --> 50:51.890]  Do you have a question?
[50:52.010 --> 50:52.430]  Yeah, I did.
[50:52.610 --> 50:53.550]  So, um...
[50:54.450 --> 50:54.770]  Um...
[50:54.770 --> 50:55.290]  Your...
[50:55.630 --> 50:55.950]  Your...
[50:56.790 --> 50:57.110]  Uh...
[50:57.110 --> 50:58.610]  Request was that...
[50:58.610 --> 50:58.750]  Uh...
[50:58.750 --> 50:59.410]  We get involved.
[51:00.710 --> 51:01.030]  Um...
[51:03.450 --> 51:03.970]  Okay.
[51:04.550 --> 51:05.630]  And the other question.
[51:05.810 --> 51:06.610]  Two more questions.
[51:06.850 --> 51:07.610]  I'll take them all together.
[51:07.950 --> 51:08.170]  Okay.
[51:08.370 --> 51:09.730]  Well, mine is in the same vein.
[51:09.950 --> 51:11.330]  I want to make sure I understand something.
[51:11.450 --> 51:23.470]  Are you saying that public officials who are involved in oversight and intelligence have actually contacted you or your colleagues with a view toward trying to understand what's going on?
[51:23.470 --> 51:31.830]  Not just to tell you that they really want you to help them understand what's going on.
[51:32.030 --> 51:32.390]  Uh...
[51:32.390 --> 51:33.070]  That's an easy question.
[51:33.150 --> 51:33.770]  The answer is yes.
[51:34.070 --> 51:34.690]  Next one.
[51:49.290 --> 51:49.770]  Uh...
[51:49.770 --> 51:53.190]  Reaching out to folks in government is surprisingly easy.
[51:53.310 --> 51:57.890]  I actually worked a little bit on the other side of the table at the California Department of Justice.
[51:58.070 --> 52:02.150]  And we loved when people volunteered information that was helpful to us.
[52:02.530 --> 52:02.890]  Um...
[52:02.890 --> 52:06.030]  And if, you know, folks came to us and said, what can we do that'll be helpful to you?
[52:06.070 --> 52:06.630]  We would tell them.
[52:06.710 --> 52:08.590]  And that's been my experience in this debate too.
[52:09.010 --> 52:09.370]  Um...
[52:09.370 --> 52:12.870]  That there are folks with big questions they want answered and they don't know how to answer them.
[52:12.870 --> 52:13.410]  Uh...
[52:13.410 --> 52:15.370]  There are also the questions they don't know how to ask.
[52:15.510 --> 52:18.310]  And there, I think, the way to start is, uh, what we tried to do.
[52:18.410 --> 52:25.590]  Start with understanding the law, then do some technical studies of what the law means, and then circle that back to, and here's why the law is probably wrong.
[52:26.750 --> 52:27.110]  Um...
[52:27.110 --> 52:27.230]  Uh...
[52:27.230 --> 52:29.730]  With respect to learning the law in this area, um...
[52:30.590 --> 52:30.950]  Uh...
[52:30.950 --> 52:33.330]  We've been frustrated with the quality of resources available.
[52:33.510 --> 52:35.890]  And that's actually one of the reasons why we're working on this online class.
[52:36.070 --> 52:40.810]  So that, uh, folks who are really interested can get a good handle for free, uh, on the law of the space.
[52:41.830 --> 52:42.190]  Um...
[52:42.190 --> 52:42.350]  Uh...
[52:42.350 --> 52:45.450]  Tempora, uh, appears to be, uh, a British program.
[52:46.110 --> 52:46.670]  Um...
[52:46.670 --> 52:46.750]  Uh...
[52:46.750 --> 52:49.110]  So it's, uh, operated by GCHQ.
[52:49.290 --> 52:50.410]  It's not under Section 702.
[52:50.890 --> 52:51.850]  It's outside the U.S.
[52:52.490 --> 52:52.910]  Um...
[52:52.910 --> 52:53.990]  So, uh...
[52:53.990 --> 52:56.810]  So the legal authority for the two programs is very different.
[52:57.610 --> 52:58.170]  Um...
[52:58.170 --> 52:58.390]  Uh...
[52:58.390 --> 52:59.490]  Collection from the...
[52:59.490 --> 53:09.970]  The Brits would, uh, seemingly fall under Executive Order 12333, which means the sum total of your privacy protection is the Fourth Amendment as interpreted by secret, in secret by the executive branch.
[53:10.610 --> 53:11.170]  Um...
[53:11.170 --> 53:17.610]  Which is, uh, you know, to the extent FAA 702 seems a little lax, uh, that may be a little laxer.
[53:19.250 --> 53:19.530]  Okay.
[53:19.790 --> 53:20.130]  Thanks, everyone.
[53:28.990 --> 53:29.270]  Um...
[53:29.270 --> 53:31.310]  I'm not sure what the deal is with, uh, focus stuff.
[53:31.490 --> 53:33.450]  If they have a place I can send them slides, I'll send them slides.
[53:38.490 --> 53:38.770]  Um...
[53:38.770 --> 53:46.790]  Do you have a website?
[53:47.270 --> 53:47.910]  Oh, sure.
[53:49.390 --> 53:49.930]  I'm sorry?
[53:50.170 --> 53:51.430]  Your website, what is it?
[53:51.450 --> 53:53.590]  Oh, it's just first name, last name, dot org.
[53:55.890 --> 53:56.750]  Uh, org?
[53:57.010 --> 53:57.270]  Yeah.
[53:57.610 --> 53:58.890]  I wish I could get .com.
[54:00.430 --> 54:00.490]  Oh.
[54:03.610 --> 54:05.570]  Quick follow-up question.
[54:06.070 --> 54:16.950]  If you need, like, a less popular video conferencing program, you know, like obscure, does that make it safer?
[54:17.510 --> 54:22.150]  Uh, so, I think safer is all relative to what your threat model is.
[54:22.150 --> 54:23.490]  Uh, if you're concerned...
[54:23.490 --> 54:23.690]  Uh, if you're concerned...