1
00:00:00,460 --> 00:00:04,960
Alright, so before we start, a little request to please keep your phones on silent.

2
00:00:05,380 --> 00:00:13,340
And there will be a Q&A round, and if there are further questions after this talk is completed, we have a speaker's lounge right next door.

3
00:00:13,340 --> 00:00:17,380
So you can meet up with the speaker, and you can have your questions addressed over there.

4
00:00:17,380 --> 00:00:23,480
Also, if you like this particular talk, please tweet us at DefConVillage...

5
00:00:23,480 --> 00:00:30,360
at ReconVillage on Twitter and on LinkedIn, and we'll be handing out special gifts and prizes for some random people.

6
00:00:30,760 --> 00:00:40,550
Alright, so the next speaker that we have is Patrick Warren, and just a little introduction about him.

7
00:00:40,550 --> 00:00:45,530
He is an Associate Professor in the Department of Economics at Clemson University.

8
00:00:45,730 --> 00:00:54,870
He co-directs the Media Forensic Hub at Clemson University, where he does a bunch of work using OSEAN techniques, mostly in the context of social media.

9
00:00:54,870 --> 00:01:08,010
And the topic that he will be covering is, Oh, the Places You Will Go Using Media Variants to Trace the Organization and Behavior of a Coordinated Inauthentic Influence Operation.

10
00:01:08,010 --> 00:01:11,010
So, yes, please welcome Patrick Warren.

11
00:01:15,820 --> 00:01:17,400
Thank you so much.

12
00:01:17,400 --> 00:01:19,720
I did not coordinate with the speaker before.

13
00:01:19,720 --> 00:01:28,440
We both randomly picked, oh, the places you'll go, but I guarantee to you this is all out of my own brain and not GPT created.

14
00:01:30,640 --> 00:01:31,720
I'm an academic.

15
00:01:31,720 --> 00:01:42,600
I'm going to tell you about a project that we've been working on at the Hub with a big team of undergrads, where we teach them some basic open source techniques and then turn them loose on a problem that we find interesting and we think they'll find interesting.

16
00:01:42,600 --> 00:01:48,280
And they often will come up with new and interesting ways to look at stuff that we wouldn't have thought of at all.

17
00:01:49,020 --> 00:01:58,300
And so, if you like this, if you find this interesting, you can go to our website, just Google Media Forensics Hub Clemson, and you can find a bunch of reports that our students have created.

18
00:01:58,640 --> 00:02:00,340
And I also want to thank Three Gimbals.

19
00:02:00,340 --> 00:02:03,100
They do a bunch of open source stuff for money.

20
00:02:04,120 --> 00:02:09,120
And they've partnered with our students quite a bit and they sponsored us to come out here, so thanks to them.

21
00:02:09,660 --> 00:02:11,520
So, pictures.

22
00:02:11,520 --> 00:02:12,640
Oh, the places you'll go.

23
00:02:12,680 --> 00:02:14,860
So, here's what we're going to do today.

24
00:02:14,860 --> 00:02:16,180
We're talking about Dragon Bridge.

25
00:02:16,180 --> 00:02:20,160
So, Dragon Bridge is a Chinese originating disinformation actor.

26
00:02:20,160 --> 00:02:23,680
You may have heard them called Spamouflage Dragon or Dragon Bridge.

27
00:02:23,680 --> 00:02:33,220
For folks who have worked in the kind of coordinated and authentic state-backed information operations space, they're probably the most prolific Chinese actor in that space.

28
00:02:33,220 --> 00:02:36,120
If you know anything about Chinese info ops, you probably know about them.

29
00:02:36,140 --> 00:02:38,220
But just in case, I'm going to tell you about them a little bit.

30
00:02:38,560 --> 00:02:44,440
And then the guts of the talk, the main point, is this idea of using media variances data.

31
00:02:44,580 --> 00:02:46,760
So, I'm going to talk a lot about what I mean by that.

32
00:02:47,080 --> 00:02:52,780
And then I'm going to compare using media to using text and why it might be better or some reasons it might be better.

33
00:02:52,780 --> 00:03:04,780
And then maybe, if we have time, I'll take the whole thing to 11 and think about rather than using media or using text as data, can we use theories as data?

34
00:03:04,980 --> 00:03:07,000
So, hopefully we'll get there.

35
00:03:08,320 --> 00:03:10,740
So, Spamouflage Dragon or Dragon Bridge.

36
00:03:10,880 --> 00:03:11,800
We have a report.

37
00:03:11,800 --> 00:03:12,560
You can Google it.

38
00:03:12,560 --> 00:03:13,660
Google the five-year span.

39
00:03:13,660 --> 00:03:17,800
This is kind of a longitudinal investigation of this actor and how they work.

40
00:03:18,480 --> 00:03:20,540
They've been around for at least five years.

41
00:03:20,540 --> 00:03:21,760
Bandian has reported on them.

42
00:03:21,760 --> 00:03:22,940
Graphic has reported on them.

43
00:03:23,060 --> 00:03:24,860
Google has talked about them a lot.

44
00:03:25,000 --> 00:03:26,580
A lot of people have talked about them.

45
00:03:26,760 --> 00:03:39,680
But I think the best way to understand them is as a persistent Chinese influence operation that uses tons and tons of accounts that are immediately obviously fake and yet still manage to have...

46
00:03:39,680 --> 00:03:41,980
And I'll argue about this during Q&A if you want.

47
00:03:41,980 --> 00:03:43,420
That's not really what this talks about.

48
00:03:43,420 --> 00:03:48,360
But still, despite having accounts that are obviously fake, can still have an enormous impact.

49
00:03:48,360 --> 00:03:49,980
And we can think about why that might be.

50
00:03:50,540 --> 00:03:51,680
That's not what this talks about.

51
00:03:51,680 --> 00:04:02,800
What this talks about is using one of the techniques that Dragon Bridge uses to have the impact that they do in order to unpack something about the way that organization operates.

52
00:04:02,980 --> 00:04:06,280
So, just in case you don't know anything about Dragon Bridge at all, here's what they like to talk about.

53
00:04:06,280 --> 00:04:08,400
They like to talk about how the USA sucks.

54
00:04:08,640 --> 00:04:15,120
They like to talk about Guo, the Chinese dissident billionaire who's now in trouble with the DOJ.

55
00:04:15,520 --> 00:04:18,140
Miles Guo, Miles Kwok, he goes by all kinds of different names.

56
00:04:18,140 --> 00:04:20,700
They hate him and have hated him for a long time.

57
00:04:21,100 --> 00:04:26,480
But then they like to talk about Xinjiang, they like to talk about Hong Kong, they like to talk about Taiwan recently.

58
00:04:26,780 --> 00:04:31,460
One of their very favorite topics are what we call problematic expat Chinese people.

59
00:04:31,480 --> 00:04:37,300
So basically, anyone who's saying bad things about the Chinese government while overseas, they hate those people.

60
00:04:38,220 --> 00:04:42,860
And how do they try to influence conversations around these topics?

61
00:04:42,860 --> 00:04:45,100
Well, they do all kinds of different stuff.

62
00:04:45,400 --> 00:04:59,740
Original cartoons, original video, multilingual recycling, hashtag flooding, calls to action, lots of baby accounts, lots of account spoofing, lots of hacked accounts, little bit of cross-platform stuff, little bit of GAN faces.

63
00:04:59,740 --> 00:05:05,860
We heard a little bit about the way you can do GAN faces better now in the sort of generative AI world.

64
00:05:06,120 --> 00:05:08,740
But we're going to concentrate on these things.

65
00:05:08,780 --> 00:05:11,780
These are techniques that they've used since their inception.

66
00:05:11,780 --> 00:05:15,700
And this is basically creating from scratch novel media.

67
00:05:16,060 --> 00:05:17,600
Now, it's not great.

68
00:05:17,840 --> 00:05:19,580
I'll show you what it looks like.

69
00:05:19,780 --> 00:05:21,380
This is what it looks like.

70
00:05:21,400 --> 00:05:23,220
It might be hard for the folks in the back.

71
00:05:23,220 --> 00:05:25,800
There's plenty of seats if you want to come up where you can see.

72
00:05:26,880 --> 00:05:39,560
It's basically really, really mediocre hand-drawn or computer-drawn cartoons or videos where they superimpose AI announcers in the background.

73
00:05:39,560 --> 00:05:42,580
Often they'll do these cartoons in lots of different languages.

74
00:05:42,580 --> 00:05:48,860
They'll just recycle the same cartoon and put the little captions in Chinese and in English and in French.

75
00:05:48,860 --> 00:05:49,640
I don't really know.

76
00:05:49,640 --> 00:05:51,780
Well, I have some ideas about what they might do in French.

77
00:05:51,880 --> 00:05:53,200
But this is what we're talking about.

78
00:05:53,200 --> 00:06:00,660
Think about hundreds and hundreds or even up to a thousand such media objects that have been created by this campaign.

79
00:06:02,460 --> 00:06:04,220
Okay, so now to the guts.

80
00:06:04,220 --> 00:06:06,960
So now you know a little bit about Dragon Bridge and what they're up to.

81
00:06:06,960 --> 00:06:14,620
I want to talk about how we might be able to use this media against them, in a sense, or certainly to learn about them.

82
00:06:15,000 --> 00:06:20,300
And so here's the process that I'm going to talk you through today, just in one slide.

83
00:06:20,300 --> 00:06:21,140
So what do we start with?

84
00:06:21,140 --> 00:06:22,940
We start with a seed set.

85
00:06:22,940 --> 00:06:31,760
We start with a couple hundred examples of cartoons that we think were created by Dragon Bridge.

86
00:06:31,760 --> 00:06:34,900
And again, they don't hide who they are really at all.

87
00:06:34,900 --> 00:06:41,620
It's pretty easy to find cartoons that you can attribute at a very high confidence to this operator.

88
00:06:41,620 --> 00:06:46,640
And so cartoons or YouTube videos are still from YouTube videos, or the whole video, we can think about that.

89
00:06:46,640 --> 00:06:54,240
But we need a little seed set of images or videos that we are pretty sure are coming from this actor.

90
00:06:55,280 --> 00:06:59,860
Then we need to turn that seed into a field.

91
00:07:00,040 --> 00:07:02,520
And the way to do that is reverse image search.

92
00:07:02,520 --> 00:07:05,100
There's lots of ways that you can automate this reverse image search.

93
00:07:05,100 --> 00:07:06,800
I won't talk through all of them.

94
00:07:07,080 --> 00:07:08,340
You can roll your own.

95
00:07:08,340 --> 00:07:10,100
There's search API ways to do this.

96
00:07:10,100 --> 00:07:11,880
You can actually pay Google to do it.

97
00:07:11,880 --> 00:07:12,820
It's not just Google.

98
00:07:12,820 --> 00:07:17,340
Of course, there's many, many reverse image search engines that you can use.

99
00:07:17,340 --> 00:07:20,840
You really want to cast just as broad of a net as you can.

100
00:07:20,920 --> 00:07:28,840
Because what we're looking for is all of the different variants of this image that pop up all over the internet.

101
00:07:28,840 --> 00:07:31,280
As many as we can find in as many places as we find.

102
00:07:31,280 --> 00:07:34,400
And I'll tell you, for this actor, you're going to find them everywhere.

103
00:07:34,580 --> 00:07:44,680
So if you pick a topic where you know this actor is active, and you search on that topic on basically any place where someone can post images on the internet, you're going to find images from this actor.

104
00:07:44,680 --> 00:08:03,100
I mean, in a Hindi website that helps Indian high school students cheat on their homework, if one of the things you can do is post the solution as an image, you're going to find a picture of Miles Guo on that website ripping off ants.

105
00:08:04,160 --> 00:08:06,440
So they are literally everywhere.

106
00:08:06,520 --> 00:08:12,200
And so once you plug these seed images into these reverse image search, you're going to get lots and lots and lots and lots of these.

107
00:08:12,200 --> 00:08:14,960
And you're going to find lots of them in social media platforms.

108
00:08:14,960 --> 00:08:17,800
And of course, every social media platform you can imagine.

109
00:08:17,800 --> 00:08:20,860
I've just shown some here, but all of them.

110
00:08:21,700 --> 00:08:23,020
Many variants.

111
00:08:23,260 --> 00:08:24,300
But why do we care?

112
00:08:24,300 --> 00:08:27,120
These are just more examples of the same thing.

113
00:08:27,340 --> 00:08:29,560
Well, there's a couple reasons we care.

114
00:08:29,560 --> 00:08:34,220
The first reason we care is that those variants can become new seeds.

115
00:08:34,300 --> 00:08:36,040
How do they become new seeds?

116
00:08:36,040 --> 00:08:45,520
Well, if I have one account that's posted ten variants of one of the seeds, and then they've posted some other image that isn't one of the seeds, well, now I've probably got a new seed.

117
00:08:45,560 --> 00:08:46,940
That's just the way these accounts work.

118
00:08:46,940 --> 00:08:55,220
So I can go from 200 seeds to 400 seeds to 800 seeds, and then for each of the seed images, I can find lots and lots and lots of variants.

119
00:08:55,320 --> 00:08:56,960
Why do I care about variants?

120
00:08:56,960 --> 00:09:03,220
Because the variants will sometimes get posted in places where the metadata remain intact.

121
00:09:03,280 --> 00:09:10,820
So you may start with images on Twitter, but from a metadata perspective, images on Twitter are pretty weak sauce, right?

122
00:09:10,820 --> 00:09:16,600
I can see the dimensions and things like that, but most of the metadata are gone by the time they show up on Twitter.

123
00:09:16,600 --> 00:09:24,920
But if instead they posted on IndianHomeworkHelp.com, and they haven't been careful enough, they may have left a trace that I can see.

124
00:09:25,360 --> 00:09:37,280
And not only might they have left a trace that I can see, I may find those same traces on some other websites, but on some other variants of that image, I see different traces, and now, look, I've learned something about the way they post their images.

125
00:09:37,280 --> 00:09:42,280
Because however they're posting them here is not the same way they're posting them there, because the traces look different in these two cases.

126
00:09:42,600 --> 00:09:44,620
So both of these things are valuable.

127
00:09:44,620 --> 00:09:54,040
Both running it through the social media profiles to find more seeds, and then extracting the metadata not just from each seed, but from this big old scope worth of variants.

128
00:09:54,980 --> 00:09:57,240
So what are the payoffs you'd get out of this?

129
00:09:57,420 --> 00:09:58,240
So let's say I do this.

130
00:09:58,240 --> 00:09:58,860
Imagine I do it.

131
00:09:58,860 --> 00:10:01,180
I've left some of the details, obviously, vague.

132
00:10:01,340 --> 00:10:02,580
What can I get out of it?

133
00:10:02,580 --> 00:10:03,620
Well, profiles.

134
00:10:03,780 --> 00:10:05,440
Profiles are actually valuable, right?

135
00:10:05,440 --> 00:10:11,960
So I start with these sets of images, and now I have images, and I also have profiles, social media profiles across lots of places.

136
00:10:11,960 --> 00:10:13,960
I might be able to get profile information.

137
00:10:13,960 --> 00:10:16,740
I may learn something about how those profiles got created.

138
00:10:16,820 --> 00:10:29,780
For example, in one of the cases that we tracked, which had to do with Xinjiang cotton, we found that a ton of the profiles that they had posted with on Twitter, they had actually set up with hacked accounts from the same hack release.

139
00:10:29,780 --> 00:10:31,060
That's useful to know.

140
00:10:32,640 --> 00:10:35,200
I may also learn something about the software that they're using.

141
00:10:35,200 --> 00:10:39,680
So here I have Photoshop, Adobe Photoshop.

142
00:10:39,840 --> 00:10:41,460
Why do I have Adobe Photoshop here?

143
00:10:41,460 --> 00:10:44,600
Common image editing software.

144
00:10:44,720 --> 00:10:50,320
And metadata about Photoshop is often left in images that get edited with Photoshop.

145
00:10:50,440 --> 00:10:54,060
That can include things like what version of Photoshop got used.

146
00:10:54,060 --> 00:10:56,480
Is it a Windows version or a Mac version?

147
00:10:56,480 --> 00:10:57,500
I might find out that.

148
00:10:57,500 --> 00:10:59,080
What year of the Photoshop was used?

149
00:10:59,080 --> 00:11:00,400
I might find out that.

150
00:11:00,400 --> 00:11:05,300
Was there different directories that were used to make that Photoshop image?

151
00:11:05,300 --> 00:11:09,880
Sometimes that gets left, and so I've learned directory structure now, at least for that variant.

152
00:11:11,300 --> 00:11:19,540
I have a signature here because some idiots put their name in the directory structure, which is crazy, but happens, and has happened here.

153
00:11:20,380 --> 00:11:23,520
Hardware specs, sometimes that gets captured in the metadata.

154
00:11:23,940 --> 00:11:30,820
And of course the Holy Grail, sometimes they leave the GPS information in the metadata, and then we're like, oh well, this was in Yulin.

155
00:11:30,820 --> 00:11:32,140
That's kind of nice to know.

156
00:11:35,600 --> 00:11:43,620
Also, I can use these individual bits of data about the variants to do a couple things.

157
00:11:43,840 --> 00:11:46,280
I can cluster them together.

158
00:11:46,680 --> 00:12:08,900
So if I see combinations of the variants being used on these sites and not these sites, or by these accounts and not these accounts, I know that the guy that's posting on IndianHomeworkHelp.com is probably the same guy that's posting on DWNews.com, because the images have identical metadata.

159
00:12:09,460 --> 00:12:20,380
And in fact, that person is posting the same set of images in both those places, but it's different than the set of images or the metadata that are getting posted on Quora, or Pinterest, or wherever.

160
00:12:20,380 --> 00:12:28,240
And so now I can cluster users or accounts, and sometimes I can even learn something about the structure.

161
00:12:28,240 --> 00:12:30,340
And what do I mean by the structure here?

162
00:12:30,340 --> 00:12:36,720
So if you know anything about Chinese information operations, you probably also know something about Russian information operations.

163
00:12:36,720 --> 00:12:41,240
If you know anything about Russian information operations, probably what you know is the Internet Research Agency.

164
00:12:41,940 --> 00:12:45,220
And they're famous for having a centralized media shop.

165
00:12:45,360 --> 00:12:46,080
What do we mean?

166
00:12:46,080 --> 00:12:52,320
We mean that they had some guys that made all their media, and then they used those media on a lot of different accounts all over the place, but it was centralized.

167
00:12:52,320 --> 00:12:57,560
So the media was created, at least back in 2016-2017, was all created in sort of the same place.

168
00:12:58,060 --> 00:13:04,240
But if I see that there are 17 different versions of Adobe being used to create these images, then they're not all happening in the same shop.

169
00:13:04,240 --> 00:13:06,100
This has been farmed out.

170
00:13:06,500 --> 00:13:14,160
And is it being farmed out to individuals, or contractors, like Fiverr contractors, or is it getting farmed out to marketing shops?

171
00:13:14,160 --> 00:13:16,360
The patterns can tell us something about it.

172
00:13:18,450 --> 00:13:19,550
So, application.

173
00:13:19,550 --> 00:13:21,010
So that's the idea of theory.

174
00:13:22,390 --> 00:13:25,370
In practice, this is what it looked like in this...

175
00:13:25,970 --> 00:13:32,370
It's called All the Places You'll Glow, because it was the title of the report, where we looked at this campaign that focused on these three guys.

176
00:13:32,610 --> 00:13:36,550
Steve Bannon, Lieberman, and Miles Guo.

177
00:13:37,790 --> 00:13:44,810
We started with 236 seed images that we found, mostly on mainstream social media, Twitter.

178
00:13:46,510 --> 00:13:52,610
Once we run them through some very simple search functions, we end up with 2,700 variants of these 236 images.

179
00:13:52,610 --> 00:13:54,250
Again, this is a huge underestimate.

180
00:13:54,250 --> 00:13:56,190
There are many, many more variants out there.

181
00:13:56,190 --> 00:13:59,730
But this is what was still on... this was just using Google or a semi-search.

182
00:14:00,470 --> 00:14:09,050
And then, once we see those variants, we can then infer some social media profiles that are probably central to this campaign, because they posted many of those variants.

183
00:14:09,190 --> 00:14:20,830
Once they posted many of those variants, I was able to extract 200 new seeds, so images that were not in my original set of images, because those images were also posted by those accounts that were central to this campaign.

184
00:14:20,970 --> 00:14:24,210
And from those 200 new seeds, I ended up with 800 new variants.

185
00:14:24,210 --> 00:14:37,070
So I ended up with, I don't know, about 3,500 variants that were all basically pictures of these three guys, Lieberman, Steve Bannon, and Miles Guo, in all kinds of different languages.

186
00:14:37,070 --> 00:14:46,290
And just to give you a sense of where these were, this is just the top-level domains where we found these images.

187
00:14:46,650 --> 00:14:48,630
And here's the order they're in.

188
00:14:48,630 --> 00:14:50,330
So Twitter, number one.

189
00:14:50,730 --> 00:14:51,730
Reddit.

190
00:14:51,910 --> 00:14:54,250
Does anybody know what this is?

191
00:14:56,560 --> 00:14:58,500
No fair if you're with the team.

192
00:15:00,580 --> 00:15:02,080
Nyonga Money.

193
00:15:02,340 --> 00:15:06,020
It's like a Nigerian crypto forum.

194
00:15:06,020 --> 00:15:09,620
It is the number three place that these things showed up.

195
00:15:10,180 --> 00:15:11,060
OK.

196
00:15:11,200 --> 00:15:13,580
Who's surprised by this one?

197
00:15:14,720 --> 00:15:20,840
Yeah, your mom is seeing a lot of pictures of Miles Guo.

198
00:15:21,860 --> 00:15:32,240
In fact, they were so successful on Pinterest that Pinterest shut down the ability to search for the term Lieberman, the Chinese virologist that they targeted here.

199
00:15:32,240 --> 00:15:34,740
You can't search for Lieberman on Pinterest right now.

200
00:15:34,740 --> 00:15:36,260
They've shut down the topic.

201
00:15:37,420 --> 00:15:44,800
I said I wasn't going to argue about this, but this is one of the ways that they're successful in ways that people don't quite appreciate.

202
00:15:44,840 --> 00:15:53,160
They can affect the topic without having accounts that convince anyone, if they can manage to flood out the target so much that people can't talk about it.

203
00:15:53,180 --> 00:15:54,920
And Pinterest has helped them do that.

204
00:15:55,520 --> 00:15:56,120
Flickr.

205
00:15:56,120 --> 00:15:56,940
I love Flickr.

206
00:15:56,940 --> 00:15:58,240
Why do I love Flickr?

207
00:15:59,020 --> 00:16:00,640
Metadata intact on Flickr.

208
00:16:00,640 --> 00:16:02,480
Metadata intact on Pinterest.

209
00:16:02,480 --> 00:16:11,360
Metadata intact on the Nigerian cryptocurrency, but not on Twitter, not on Reddit, by default, anyway.

210
00:16:11,840 --> 00:16:16,680
OK, so there's a lot, 180 of these, that we found substantial numbers of these images.

211
00:16:19,260 --> 00:16:22,100
Substantial variance in the software being used to post these images.

212
00:16:22,100 --> 00:16:23,920
I've sort of previewed that already.

213
00:16:24,000 --> 00:16:28,660
I mean, Adobe Photoshop 2020 was the number one, but there's things I had never heard.

214
00:16:28,660 --> 00:16:30,040
PicsArt, I didn't know PicsArt.

215
00:16:30,060 --> 00:16:34,080
MeToo is a Chinese, I think, web-based photo editing thing.

216
00:16:34,080 --> 00:16:39,440
There was MeToo, but tons of variance, including tons of different systems.

217
00:16:39,440 --> 00:16:42,880
So, Windows, cloud-based, Mac.

218
00:16:42,880 --> 00:16:47,220
This was not, not, not, not a centralized art shop, for sure.

219
00:16:48,700 --> 00:16:51,660
Unless you're having everybody bring in their system to do it, right?

220
00:16:51,660 --> 00:16:53,080
Very different structure.

221
00:16:55,420 --> 00:16:59,040
Multigo, had to use Multigo a little bit for the OSINT class.

222
00:16:59,060 --> 00:17:04,340
This is just a little, when I was talking about clustering them together, this is just an illustration of that.

223
00:17:04,500 --> 00:17:12,960
I can connect images by devices, or images by software, or images by accounts, and you can get fun little clusters.

224
00:17:12,960 --> 00:17:26,100
And sometimes, those clusters include, as I mentioned, our kind of fun gold standard, which is, I could find a bunch of images that were all posted, and oops, a couple times, they forgot to clean the GPS off them, and that's where they were.

225
00:17:26,920 --> 00:17:28,980
I could show you the house they were in.

226
00:17:29,060 --> 00:17:31,440
It was not in an industrial area.

227
00:17:31,600 --> 00:17:33,700
This had been outsourced, for sure.

228
00:17:35,020 --> 00:17:46,300
This is an example in a different context, but I was talking about the file path, so we did some work looking at some metadata on videos released around the Russian invasion of Ukraine.

229
00:17:46,300 --> 00:17:49,120
These are, like, fact checks that had been faked.

230
00:17:49,520 --> 00:18:05,300
So these were, like, imagine that you saw a fact check that had a burning Russian piece of machinery, and then a picture of a non-burning piece of machinery, and they said, look, this has been photoshopped by the Ukrainians to try to convince you that they're blowing up our stuff,

231
00:18:05,300 --> 00:18:06,540
but it's not true.

232
00:18:06,640 --> 00:18:13,800
But then you dig in the metadata, and you see that, oh, look, whoever created this fact check also created the burning image.

233
00:18:13,800 --> 00:18:19,760
So they have created fake news and then fact-checked it to convince you that the Ukrainians are lying.

234
00:18:20,040 --> 00:18:28,720
But they forgot to clean up their metadata, and it turns out it was in a directory called Gerasimov-Constantin Working Folder.

235
00:18:29,620 --> 00:18:40,880
Probably not every variant of that media had this metadata still intact, but if you search hard and far and long enough, you might find a variant that does.

236
00:18:40,880 --> 00:18:43,780
So just because you found it once doesn't mean you're done.

237
00:18:43,840 --> 00:18:48,900
You want to find as many variants as you can, because it just takes one slip.

238
00:18:52,310 --> 00:18:54,590
OK, could we do all this with text?

239
00:18:54,590 --> 00:18:58,330
So instead of media, instead of a book with text, could we do all this?

240
00:18:58,330 --> 00:19:00,090
I think we could do some of this, still.

241
00:19:00,210 --> 00:19:02,010
This process isn't crazy.

242
00:19:02,290 --> 00:19:06,690
You could start with some text, and these actors use long text, too.

243
00:19:06,990 --> 00:19:18,570
Duplicate those texts, search for those texts, find other accounts that use the exact same text, pull them together, identify new texts by looking at the text that accounts that post a lot of these seed texts also post.

244
00:19:18,610 --> 00:19:22,650
You could do this, but the metadata is not going to work very well.

245
00:19:22,650 --> 00:19:31,110
I mean, I guess you could use font or something, specific word choices, maybe, but metadata, I mean, they're just going to exactly duplicate.

246
00:19:31,410 --> 00:19:32,950
It's not as obvious.

247
00:19:32,950 --> 00:19:43,910
And the other thing is that the way people use social media tends to differ a lot by platform.

248
00:19:43,910 --> 00:19:50,050
So it would be rare, for instance, that the screed that I point on Reddit, I would post on Twitter.

249
00:19:50,050 --> 00:19:51,730
It won't fit on Twitter.

250
00:19:52,050 --> 00:19:59,690
And so, people reuse images, at least these guys, and I think many people who work in this domain, reuse images in a way that they don't reuse text.

251
00:19:59,690 --> 00:20:01,830
And so, you can't do this for text.

252
00:20:01,830 --> 00:20:03,150
It doesn't work quite as well.

253
00:20:04,410 --> 00:20:07,650
But here's the problem, or a problem with this approach.

254
00:20:07,890 --> 00:20:09,990
This works because media is expensive.

255
00:20:11,050 --> 00:20:13,730
I reuse media because it's expensive.

256
00:20:13,730 --> 00:20:15,990
I've got to hire somebody else to draw a new picture.

257
00:20:16,610 --> 00:20:18,910
But you know, let's go back to the talk two times ago.

258
00:20:19,450 --> 00:20:21,810
Media is not as expensive as it used to be.

259
00:20:21,850 --> 00:20:47,690
This guy said, if I were an artist, I'd be scared to death, because my bad-ass tiger here was created with a generative AI, and I could create three tigers with turtlenecks tomorrow, and I could create thousands of tigers with turtlenecks tomorrow, and reverse image search isn't going to tie those tigers with turtlenecks together,

260
00:20:47,690 --> 00:20:55,350
probably, at least not traditional reverse image search, because they don't look enough alike to kind of get matched up by reverse image search.

261
00:20:55,450 --> 00:21:07,430
And so, although these are clearly all variants of the theme of tigers with turtlenecks, less bad-ass than the neon tiger, of course, it's hard to reverse image search for a theme.

262
00:21:08,210 --> 00:21:19,610
And so, when I said that maybe we could take this to 11, I think there is still a path forward for this idea of using media variants, but that path is going to use AI to fight AI.

263
00:21:19,610 --> 00:21:33,070
So, maybe reverse image search can't say those are all tigers with turtlenecks, but computer vision can say those are all tigers with turtlenecks, or at least it can get pretty close to saying those are all tigers with turtlenecks.

264
00:21:35,330 --> 00:21:41,870
I've been talking about pictures here, it's not just pictures, there's videos, too, but we heard about Whisper this morning.

265
00:21:42,470 --> 00:21:44,330
We can do this with images, too, right?

266
00:21:44,330 --> 00:21:53,030
You combine computer vision with some screenshots out of a video, and now we've got the transcript, and at least we know at least every 10 seconds what was on the screen.

267
00:21:53,090 --> 00:22:04,130
We can get a pretty good sense of what the video is about, and so a video about a tiger and the turtleneck, we might be able to match it to other videos about tigers and turtlenecks, or about how Taiwan is really part of China.

268
00:22:05,370 --> 00:22:14,610
And so, we may be able to find variants of videos or media in the same way that we find variants of exact images.

269
00:22:14,610 --> 00:22:16,230
It's just going to be a little harder.

270
00:22:17,750 --> 00:22:25,190
The idea I guess I want to leave you with here is that maybe tigers with turtlenecks is not actually what we want to match on.

271
00:22:25,190 --> 00:22:32,390
What we want to match on is cute animals with turtlenecks, or something.

272
00:22:32,590 --> 00:22:40,090
What we really want to do is we want to stick media items together in the same way that the influence actor would.

273
00:22:40,090 --> 00:22:56,510
And so, if I could guess or have some sense of what sort of target they might have, then I could use this same process where I look for images or videos or whatever that fit that target, and run the exact same idea.

274
00:22:56,510 --> 00:23:01,470
Or, if I have seeds of accounts, I can extract their target.

275
00:23:01,470 --> 00:23:05,230
Not at the level of tigers with turtlenecks, but at a more general level.

276
00:23:05,890 --> 00:23:06,590
This is hard.

277
00:23:06,590 --> 00:23:07,690
This is really hard.

278
00:23:07,690 --> 00:23:13,350
We're trying to teach GPT-4 right now to extract what you might call narratives.

279
00:23:13,350 --> 00:23:15,990
We like to call theories, because we think it's more general than narratives.

280
00:23:16,250 --> 00:23:19,210
You might be able to extract theories from tweets.

281
00:23:19,210 --> 00:23:22,110
You might be able to extract theories from images and videos.

282
00:23:22,170 --> 00:23:24,630
And then you can run all of this at the level of the theory.

283
00:23:24,630 --> 00:23:28,650
It's pretty abstract, but it may be where we need to go.

284
00:23:29,610 --> 00:23:33,710
I'm going to leave it there, but I'm happy to field questions about this, or about the Hub, or anything.

285
00:23:33,910 --> 00:23:35,390
And I don't know how you run questions.

286
00:23:35,390 --> 00:23:36,490
Just stand up and talk?

287
00:23:36,490 --> 00:23:37,910
Yeah, please.

288
00:24:01,700 --> 00:24:16,680
You could imagine that if I'm afraid of this sort of move by the detector, then I take my account and I seed it every once in a while with images that are purposely made to lead the LLM down the wrong track.

289
00:24:17,320 --> 00:24:19,240
That's a risk, for sure.

290
00:24:19,420 --> 00:24:22,100
That's a risk even in the approach that I'm using right now.

291
00:24:22,100 --> 00:24:33,640
If they sprinkle throughout the social media accounts some images that aren't part of the campaign at all, then it's hard for me... I'm going to match on those.

292
00:24:33,980 --> 00:24:42,220
My solution right now is, in order to consider it part of a campaign, it's going to have to be used by multiple accounts that are part of the campaign.

293
00:24:42,220 --> 00:24:44,480
Now, they may do that.

294
00:24:44,520 --> 00:24:47,860
They may stick false flag images throughout multiple accounts.

295
00:24:47,860 --> 00:24:52,140
But then, is that any different than it being part of the campaign?

296
00:24:52,200 --> 00:24:57,900
I mean, so now I've found that probably this dude who's running these ten accounts, he uses the same false flag images.

297
00:24:57,900 --> 00:24:59,140
Now I've got a trace.

298
00:24:59,140 --> 00:25:05,900
It's not a trace that's related to the product of the campaign, but rather it's a trace related to the countermeasures he's taking.

299
00:25:05,900 --> 00:25:07,320
But whatever, that's good.

300
00:25:07,320 --> 00:25:08,600
I'll count that as a trace.

301
00:25:08,600 --> 00:25:14,440
So I don't think that's quite as worrisome as you might be worried.

302
00:25:18,720 --> 00:25:20,840
Well, catch me in the speaker's lounge if you have any more questions.

303
00:25:20,840 --> 00:25:21,600
I'm happy to talk.

304
00:25:21,620 --> 00:25:22,120
Thank you.