1
00:00:02,160 --> 00:00:09,420
Thank you guys. Good morning and welcome to my talk or actually a quick guide on how to become

2
00:00:09,420 --> 00:00:17,120
your company's CEO with little effort. So first I must start with a disclaimer. So I believe the

3
00:00:17,120 --> 00:00:23,120
tech I'm presenting today is a game changer when it comes to social engineering attacks and it may

4
00:00:23,120 --> 00:00:29,660
cause harm very fast. So please remember that you can always be at the other side of the barrel

5
00:00:29,660 --> 00:00:38,340
please don't be evil, don't use deepfakes without the approval of your target. Okay, no seriously,

6
00:00:39,400 --> 00:00:45,780
but this is my target. So I'd like you to meet my company's former CEO and current chairman Udi

7
00:00:45,780 --> 00:00:53,860
Mukadi and this is him on a podcast interview he did last year. Absolutely and great to talk to you

8
00:00:53,860 --> 00:00:58,480
here. I think the beauty of the Cyborg story to fast-forward is we pioneered the space but

9
00:00:58,480 --> 00:00:59,380
continued to innovate.

10
00:00:59,660 --> 00:01:06,000
And market-leading a lot of companies don't. Yeah, so Udi was my target when I played around with deepfakes.

11
00:01:06,780 --> 00:01:13,620
And now I'd like to go over the obligatory who am I? So my name is Gal and I'm a vulnerability research

12
00:01:13,960 --> 00:01:21,960
manager at Cyborg labs, and actually my favorite research fields are embedded devices such as ISP

13
00:01:21,960 --> 00:01:23,800
equipment and mesh Wi-Fi.

14
00:01:24,840 --> 00:01:27,840
And more importantly who am I not?

15
00:01:28,240 --> 00:01:29,640
So I'm not the only one who is the target here. I am the only one. I'm the only one who is the target here.

16
00:01:29,640 --> 00:01:36,940
a machine learning expert whatsoever. This project was started as a challenge to see how far I

17
00:01:36,940 --> 00:01:43,500
can go with real time avatar with creating a real time avatar of my company's chairman. And I'm

18
00:01:43,500 --> 00:01:49,300
not going to use the A word because I bet you are all sick and tired from hearing about this

19
00:01:49,300 --> 00:01:59,520
buzzword and robot apocalypse. And as you all see I am definitely not my company's chairman. But

20
00:01:59,520 --> 00:02:08,240
this is me with a button shirt. And this is me with Udi's office as a background. And this is me

21
00:02:08,240 --> 00:02:18,380
with Udi's face. Hi, I'm Udi Sloan. And I'm very excited to be here. And it's just great to talk to

22
00:02:18,380 --> 00:02:27,980
you all. Yeah, so you probably figured out that this video was made with real time video and audio

23
00:02:27,980 --> 00:02:28,700
deep fake.

24
00:02:31,280 --> 00:02:38,940
Okay. So, for today's agenda, I would like to go over the bare minimum terms about machine

25
00:02:38,940 --> 00:02:46,320
learning. Going over the process of creating a real time video and audio deep fakes. And lastly,

26
00:02:46,400 --> 00:02:51,700
combine it all together and see how this tech can be weaponized for social engineering attack.

27
00:02:53,500 --> 00:02:58,980
Okay. So, unfortunately, I don't have enough time today to go over or be too technical about

28
00:02:58,980 --> 00:02:59,500
how this tech can be weaponized for social engineering attack.

29
00:02:59,520 --> 00:03:03,020
So I am going to tell you what I do and how this work. But I want to clear some terms before I begin.

30
00:03:03,960 --> 00:03:08,600
So, machine learning one‑on‑one, and actually, when I use the term machine

31
00:03:08,600 --> 00:03:16,060
learning or ML I actually refer to convolutional neural network or CNN which is a class in ML.

32
00:03:16,340 --> 00:03:24,420
But for the sake of simplicity, let's just call it ML. So ML model is a program that has been

33
00:03:24,420 --> 00:03:29,240
trained to recognize certain patterns and types. The most classic example,

34
00:03:29,520 --> 00:03:37,380
is image or audio recognition, and you need to train a model over a set of data providing

35
00:03:37,380 --> 00:03:43,260
it an algorithm it can use to reason over and learn from those data.

36
00:03:43,760 --> 00:03:51,720
For example, video frames and audio samples as a data set, and image or audio mapping

37
00:03:51,720 --> 00:03:52,620
as the algorithm.

38
00:03:53,920 --> 00:03:58,680
And once you have trained your model, you can use it to reason over data that it hasn't

39
00:03:58,680 --> 00:04:01,940
seen before and make prediction about those data.

40
00:04:02,140 --> 00:04:06,120
For example, detecting a face in a frame and swapping it.

41
00:04:06,660 --> 00:04:12,360
Training is basically doing a lot of mathematical calculation which takes time and often measured

42
00:04:12,360 --> 00:04:17,700
by iteration, which gives you an indication on how fast you're training your model.

43
00:04:18,100 --> 00:04:22,600
And again, please keep in mind this is a very high level explanation.

44
00:04:22,900 --> 00:04:25,640
I removed some terms just to make things simple.

45
00:04:27,500 --> 00:04:28,140
Okay.

46
00:04:28,140 --> 00:04:33,880
So, since training a model involves in a lot of math calculation, you need the right hardware

47
00:04:33,880 --> 00:04:35,440
to do it in a reasonable time.

48
00:04:35,780 --> 00:04:39,920
So GPUs were actually designed to do a lot of this kind of calculation.

49
00:04:40,260 --> 00:04:44,380
And if you've got a powerful gaming PC, you can use it to train a model.

50
00:04:45,020 --> 00:04:50,220
But if you don't, you can pay around 50 bucks per week and get it done on the cloud.

51
00:04:50,680 --> 00:04:56,220
And also, you can use Google CoLab, which is a Jupyter notebook that can actually be

52
00:04:56,220 --> 00:04:57,860
used to train a model for free.

53
00:05:00,220 --> 00:05:00,740
Okay.

54
00:05:00,840 --> 00:05:01,460
So, great.

55
00:05:01,600 --> 00:05:04,360
Now, we can talk about video and audio deepfakes.

56
00:05:05,560 --> 00:05:11,480
So the first thing you want to do when creating a deepfake avatar is gather a high-quality

57
00:05:11,480 --> 00:05:12,860
data of your target.

58
00:05:13,500 --> 00:05:17,360
This can later be used as a data set to train your model.

59
00:05:18,540 --> 00:05:22,960
So for video, I find YouTube and Google videos to be a great place to start.

60
00:05:23,180 --> 00:05:27,200
But you can also find high-quality videos in the company's website.

61
00:05:27,200 --> 00:05:28,800
Or in virtual meetings.

62
00:05:30,300 --> 00:05:34,740
For audio, here you aim for high-quality audio.

63
00:05:35,080 --> 00:05:40,340
Usually podcasts and media interviews has the best sound production.

64
00:05:40,600 --> 00:05:42,120
And they are the best source.

65
00:05:42,520 --> 00:05:49,480
But if your target is a CEO of a traded company, then you're most likely to find audio source

66
00:05:49,480 --> 00:05:50,660
in post-earning calls.

67
00:05:52,600 --> 00:05:56,320
And lastly, you might want a background.

68
00:05:56,320 --> 00:05:58,200
Which is not a data set.

69
00:05:58,400 --> 00:06:00,120
It's just a prop for staging.

70
00:06:00,400 --> 00:06:04,840
But if you want to tell a certain story, like, for example, you got stuck outside of your

71
00:06:04,840 --> 00:06:11,720
office or something like that, then you can use Google Maps or YouTube to get the right

72
00:06:11,720 --> 00:06:12,160
background.

73
00:06:12,360 --> 00:06:14,900
I used Udi's office background.

74
00:06:15,020 --> 00:06:18,520
I got Udi's office background from a media interview.

75
00:06:20,280 --> 00:06:20,800
Okay.

76
00:06:20,920 --> 00:06:25,280
So now we're ready to talk on how video deepfakes are done.

77
00:06:26,420 --> 00:06:35,800
So, in 2008, a talented guy called Ipirov created a project called DeepFaceLab or DFL

78
00:06:35,800 --> 00:06:38,300
for offline video deepfakes.

79
00:06:38,440 --> 00:06:43,980
And I bet you probably saw a lot of meme using this kind of project.

80
00:06:44,220 --> 00:06:49,340
And starting using DFL might be a little bit intimidating.

81
00:06:49,760 --> 00:06:51,940
You need to run different scripts.

82
00:06:52,040 --> 00:06:55,900
And there are around 32 different parameters that you need to configure.

83
00:06:55,900 --> 00:06:57,520
Why you train your model.

84
00:06:57,700 --> 00:06:59,340
And it affects how you train your model.

85
00:07:00,060 --> 00:07:04,760
But this project has an active community called Mr. DeepFakes.

86
00:07:04,900 --> 00:07:07,320
And you can get a lot of help from there.

87
00:07:09,420 --> 00:07:09,940
Okay.

88
00:07:10,000 --> 00:07:14,940
So to understand realtime deepfakes, we must understand offline deepfake first.

89
00:07:16,480 --> 00:07:21,000
So let's say I want to swap the face of the upper character with the bottom character.

90
00:07:22,340 --> 00:07:25,460
For that, I need to give DFL a source video.

91
00:07:25,900 --> 00:07:27,440
The face I wish to replace.

92
00:07:29,620 --> 00:07:34,380
And then a destination video of the target you wish to replace with.

93
00:07:34,740 --> 00:07:40,620
For good result, your dataset should include a variety of videos of your target with different

94
00:07:40,620 --> 00:07:41,920
angle and different lighting.

95
00:07:43,940 --> 00:07:46,760
And then DFL can create and train a model.

96
00:07:47,040 --> 00:07:49,860
And depends on your hardware, it should take a few days.

97
00:07:51,760 --> 00:07:55,680
And in some point in time, your model should reach enough iteration.

98
00:07:55,680 --> 00:07:58,420
To perform the face swap correctly.

99
00:08:00,320 --> 00:08:05,980
So this model is then applied on the source video to change the source face with the destination

100
00:08:05,980 --> 00:08:06,340
face.

101
00:08:07,860 --> 00:08:12,340
So in high level, DFL split the video into frames.

102
00:08:13,600 --> 00:08:17,100
Then for every frame, it apply a two stage model.

103
00:08:18,580 --> 00:08:21,820
The first stage of the model is detecting the source face.

104
00:08:23,420 --> 00:08:25,540
And the second is swapping.

105
00:08:25,540 --> 00:08:27,440
Swapping the source face with the destination face.

106
00:08:29,360 --> 00:08:34,020
After it's done for every frame, you got yourself a deepfake video.

107
00:08:36,060 --> 00:08:36,960
All right.

108
00:08:37,160 --> 00:08:43,640
And in 2021, IPRov created a new project called DeepFace Live.

109
00:08:44,000 --> 00:08:48,600
And this software can run DFL models in real time.

110
00:08:48,800 --> 00:08:51,040
And it's very straightforward to use.

111
00:08:53,000 --> 00:08:53,720
Okay.

112
00:08:53,780 --> 00:08:54,080
So.

113
00:08:54,080 --> 00:08:57,800
Now let's talk about how to train a real time DFL model.

114
00:08:59,560 --> 00:09:00,660
Same as before.

115
00:09:00,800 --> 00:09:02,420
You need a target data set.

116
00:09:02,520 --> 00:09:05,120
Again with variety of angles and lighting.

117
00:09:06,980 --> 00:09:09,240
You feed it to the same DFL project.

118
00:09:11,180 --> 00:09:17,720
But this time, since your model aimed to work on any given face, your source data sets will

119
00:09:17,720 --> 00:09:20,720
need to be different pictures of different people.

120
00:09:21,000 --> 00:09:21,100
Okay.

121
00:09:23,060 --> 00:09:23,700
Okay.

122
00:09:24,080 --> 00:09:26,780
So you need a target data set with around 10K of pre-mapped faces.

123
00:09:29,240 --> 00:09:32,100
And then you let DFL walk its magic again.

124
00:09:32,280 --> 00:09:37,740
And after a few days, again, depends on your hardware, you get yourself a real time model.

125
00:09:39,420 --> 00:09:39,940
Okay.

126
00:09:39,980 --> 00:09:44,880
Now to put it in use, you should connect your web camera to the DeepFace Live program.

127
00:09:46,400 --> 00:09:51,320
The DeepFace Live will load the model that you created and run it on the fly for every

128
00:09:51,320 --> 00:09:52,360
frame that it gets.

129
00:09:53,540 --> 00:09:54,060
And then you can start the process.

130
00:09:54,060 --> 00:09:58,080
And then with OBS, you can direct it to everywhere, anywhere you like.

131
00:09:58,560 --> 00:10:01,040
Like for example, to do a video conference.

132
00:10:02,640 --> 00:10:04,400
But wait, there's more.

133
00:10:06,220 --> 00:10:12,780
I actually realized that by using Snap Camera, I can apply additional effect to help me improve

134
00:10:12,780 --> 00:10:13,200
cosmetic.

135
00:10:13,420 --> 00:10:15,560
For example, beard and hairstyle.

136
00:10:16,020 --> 00:10:21,700
And actually, this is how I managed not to shave my head when I did Udi's DeepFake.

137
00:10:24,060 --> 00:10:24,500
Okay.

138
00:10:25,120 --> 00:10:26,180
Let's see again.

139
00:10:26,240 --> 00:10:26,920
Okay.

140
00:10:26,920 --> 00:10:29,420
Lastly, real time DeepFakes currently has some limitations.

141
00:10:30,040 --> 00:10:37,220
Face expressions, for example, like pulling your tongue out, may not look that good.

142
00:10:37,380 --> 00:10:44,200
And also objects that covers your face like mask, hand, they can make you look unreal.

143
00:10:44,580 --> 00:10:51,320
And lastly, comparing to offline DeepFake, here you need a GPU at your end points.

144
00:10:51,480 --> 00:10:54,040
And the more powerful GPU you have, the better.

145
00:10:54,060 --> 00:11:00,880
more frames per second you're gonna get okay so now let's talk about audio deep

146
00:11:00,880 --> 00:11:08,120
fix and I'll start with spec speech to text first okay so this tech is very

147
00:11:08,120 --> 00:11:11,620
popular and today you can see many companies providing these kind of

148
00:11:11,620 --> 00:11:16,760
services quite impressively in some cases you only need a one minute of

149
00:11:16,760 --> 00:11:22,940
clear audio sample and full disclosure I paid a few dollars to 11 labs and I work

150
00:11:22,940 --> 00:11:27,140
with them but I believe that you can also get a pretty good result with an

151
00:11:27,140 --> 00:11:33,660
open-source project called cookie and after you train your model it takes text

152
00:11:33,660 --> 00:11:40,040
and translate it into speech that sounds like your target so as you might think

153
00:11:40,620 --> 00:11:44,500
text-to-speech is not enough for real-time deepfake since it's text

154
00:11:44,500 --> 00:11:51,800
based and not voice based so now for a voice based solution which is called

155
00:11:51,800 --> 00:11:52,580
voice conversion

156
00:11:52,580 --> 00:12:01,820
or VC so on March 2023 our VC project was created by Fumiyama

157
00:12:01,820 --> 00:12:10,980
FTPS and Deepomb maybe three anim enthusiastic developers and this

158
00:12:10,980 --> 00:12:16,040
project converts voice samples and support just about any language it also

159
00:12:16,040 --> 00:12:21,740
supports singing which is cool and configuring this project takes some time

160
00:12:21,740 --> 00:12:22,560
but it's worth it

161
00:12:22,580 --> 00:12:29,740
way less intimidating than DFL all right so for our VC you need

162
00:12:29,740 --> 00:12:36,840
at least 10 minutes of clear audio data set the more the better and after a few

163
00:12:36,840 --> 00:12:44,760
hours our VC created VC model this model can take any voice as an input and

164
00:12:44,760 --> 00:12:52,560
converted to the target's voice but wait there's more remember that we talked about voice Matematics.com.

165
00:12:52,580 --> 00:12:59,880
that speech text-to-speech only need one minute of audio so you can actually use

166
00:12:59,880 --> 00:13:05,200
test-to-speech model to generate a bigger data set for RBC so basically you

167
00:13:05,200 --> 00:13:13,320
only need one minute of audio for RBC as well okay and now for running the

168
00:13:13,320 --> 00:13:19,520
model on the fly so in August 22 a guide called Okada created a project called

169
00:13:19,520 --> 00:13:26,160
voice changer it supports RBC model as well as other VC models and it's very

170
00:13:26,160 --> 00:13:33,260
very straightforward to work with and so you can take your microphone connect it

171
00:13:33,260 --> 00:13:40,120
to voice changer with your RBC model and then use a program called a be audio

172
00:13:40,120 --> 00:13:47,200
that redirect the audio to your favorite program and then you can call someone

173
00:13:47,200 --> 00:13:48,960
using an voice

174
00:13:51,480 --> 00:13:57,520
deep fake and RBC model also has some limitation mainly around constant

175
00:13:57,520 --> 00:14:03,980
syllables for example if you say for a period of time it may sound unnatural

176
00:14:03,980 --> 00:14:10,060
and also voice changer has to keep a buffer of your voice and to convert it

177
00:14:10,060 --> 00:14:16,000
on the fly and that might create latency and same as deep fake live you also need

178
00:14:16,000 --> 00:14:18,540
a relatively powerful GPU at your

179
00:14:18,540 --> 00:14:18,940
and that might create latency and same as deep fake live you also need a relatively powerful GPU at your

180
00:14:18,960 --> 00:14:27,680
endpoint okay lastly I want to talk about the obvious profit reason I

181
00:14:27,680 --> 00:14:31,720
believe that real-time deepfakes is will dramatically change how social

182
00:14:31,720 --> 00:14:37,820
engineering attack are executed no more text no more emails you can expect your

183
00:14:37,820 --> 00:14:43,460
supervisor calling you and asking you to do stuff and also that of course can

184
00:14:43,460 --> 00:14:48,880
cause business damage money transfer job termination granting access you name

185
00:14:48,880 --> 00:14:53,760
it and of course it can also lead to data leakage such as confidential

186
00:14:53,760 --> 00:15:00,660
documents or credential leakage and so on but it can also create business

187
00:15:00,660 --> 00:15:05,440
fatigue think of it if you need let's say perform multi-factor authentication

188
00:15:05,440 --> 00:15:10,740
by any kind for every phone call that you do that can affect the entire

189
00:15:10,740 --> 00:15:17,680
business operation okay and actually to use the stack for social engineering

190
00:15:17,680 --> 00:15:18,000
attack

191
00:15:18,000 --> 00:15:26,760
you need to collect a data and stage a specific story and then you probably want

192
00:15:26,760 --> 00:15:31,400
to automate everything just that you will able to train the model in a

193
00:15:31,400 --> 00:15:38,840
relatively fast way and also you can use bandwidth restriction if your

194
00:15:38,840 --> 00:15:43,740
endpoint or if your model is not mature enough and you just want to get away

195
00:15:43,740 --> 00:15:45,960
with low quality or with latency issues

196
00:15:48,280 --> 00:15:56,080
okay so for conclusion since this is a social engineering game-changer and sorry

197
00:15:56,080 --> 00:16:00,400
this is I think a social engineering game-changer training a model is a

198
00:16:00,400 --> 00:16:04,100
little difficult and it takes some times and required some reading to do

199
00:16:04,100 --> 00:16:09,980
beforehand however running the model itself on the endpoint is super easy and

200
00:16:09,980 --> 00:16:15,600
super easy straightforward to use so since training a model might be high effort

201
00:16:15,600 --> 00:16:15,940
so since training a model might be high effort

202
00:16:15,960 --> 00:16:22,320
I came out with something called DEF CON video art which stands for deep fake

203
00:16:22,320 --> 00:16:30,680
conversion or for video and audio in real time or a fancy bass trip that

204
00:16:30,680 --> 00:16:37,160
automates everything that I show today so it uses pre-trained models to save

205
00:16:37,160 --> 00:16:43,620
time it using back best practice configuration I managed to find and work

206
00:16:43,620 --> 00:16:45,940
for me and I packed it all in a single day and I think that's pretty cool and

207
00:16:45,940 --> 00:16:52,220
in the all day environment dependencies so I can actually use it on my machine or

208
00:16:52,220 --> 00:16:59,040
in the cloud and hopefully in KOD a one day okay so to use it I just need three

209
00:16:59,040 --> 00:17:05,380
minutes of video and one minute of audio I then use DFL and RVC to train the

210
00:17:05,380 --> 00:17:14,040
model and after around one day for video I end up with reasonable results and

211
00:17:14,040 --> 00:17:15,920
after 30 minutes of training I got a lot better results so this is something that is really good so to be fair I have an all day in my home I should have been doing this.

212
00:17:15,920 --> 00:17:19,800
of audio, I mean, after three minutes,

213
00:17:19,840 --> 00:17:21,760
I got reasonable results for audio.

214
00:17:22,140 --> 00:17:23,940
Again, it depends on your hardware.

215
00:17:24,620 --> 00:17:28,660
And you end up with two models of audio and video.

216
00:17:30,400 --> 00:17:33,920
Okay, and then you can use it in the streaming setup

217
00:17:33,920 --> 00:17:36,600
I just demonstrated, and use it as you wish.

218
00:17:37,520 --> 00:17:39,700
And now, for a live demo.

219
00:17:40,840 --> 00:17:41,380
Okay.

220
00:17:46,020 --> 00:17:48,000
Let me just get it out of the way.

221
00:17:49,380 --> 00:17:49,820
Okay.

222
00:17:56,220 --> 00:17:57,640
Excuse me guys, just a minute.

223
00:18:04,380 --> 00:18:05,460
Maybe this is right.

224
00:18:26,620 --> 00:18:29,800
Hi, I'm Jeff Maas, founder of DEF CON.

225
00:18:29,960 --> 00:18:33,320
I hope you enjoyed this talk, because right after it,

226
00:18:33,420 --> 00:18:35,920
DEF CON is officially canceled.

227
00:18:47,360 --> 00:18:47,960
Thanks.

228
00:18:50,820 --> 00:18:52,920
Okay, so what to do?

229
00:18:53,120 --> 00:18:56,020
So, like every social engineering attack,

230
00:18:56,340 --> 00:18:57,780
awareness is a key factor.

231
00:18:58,060 --> 00:19:00,880
Today, everybody know not to trust suspicious emails

232
00:19:00,880 --> 00:19:04,160
and texts, and seems like we already reached a point

233
00:19:04,160 --> 00:19:07,360
where you must suspect a suspicious phone

234
00:19:07,360 --> 00:19:08,860
and video calls as well.

235
00:19:09,540 --> 00:19:11,680
If something is fishy, then it probably is.

236
00:19:11,940 --> 00:19:15,160
And you should always ask for additional information.

237
00:19:15,840 --> 00:19:18,720
Preferably, this kind of information that only you

238
00:19:18,720 --> 00:19:22,560
and the person at the other side of the line knows.

239
00:19:23,240 --> 00:19:27,540
Also, you can use current real-time models limitation,

240
00:19:27,920 --> 00:19:30,800
as I mentioned before, but please keep in mind

241
00:19:30,800 --> 00:19:33,800
that this limitation might be solved in the future.

242
00:19:35,420 --> 00:19:38,440
Okay, and as for technological mitigation,

243
00:19:38,960 --> 00:19:41,780
well, you can ask for multi-factor authentication.

244
00:19:42,940 --> 00:19:45,080
There are some research papers,

245
00:19:45,080 --> 00:19:49,820
and GitHub project that aim to detect deepfakes using ML,

246
00:19:50,440 --> 00:19:52,320
and there are also commercial products,

247
00:19:52,500 --> 00:19:54,440
such as Intel Bloodflow.

248
00:19:55,060 --> 00:19:58,500
And lastly, you can use other phishing mitigation,

249
00:19:58,800 --> 00:20:02,800
such as notifications and alerts that will be triggered

250
00:20:02,800 --> 00:20:05,560
when you're under a deepfake attack, something like that.

251
00:20:06,200 --> 00:20:10,800
Okay, as for future work, I plan to try to do

252
00:20:10,800 --> 00:20:14,120
a full head deepfake, including the hair and everything.

253
00:20:14,360 --> 00:20:15,060
Okay, as for future work, I plan to try to do a full head deepfake, including the hair and everything.

254
00:20:15,060 --> 00:20:18,480
I want to try to play with different source data set,

255
00:20:18,580 --> 00:20:21,020
example, just my own data set and see how it goes.

256
00:20:21,260 --> 00:20:24,780
And lastly, Colab automation or integration

257
00:20:24,780 --> 00:20:27,520
for the environment I demonstrated before.

258
00:20:28,240 --> 00:20:31,300
Lastly, I want to say thanks to Mark

259
00:20:31,300 --> 00:20:32,440
that helped me with the research

260
00:20:32,440 --> 00:20:34,460
and great tutorials I find online.

261
00:20:34,760 --> 00:20:38,700
My targets that approved the use of their image.

262
00:20:38,880 --> 00:20:40,420
And lastly, the Archer TV show,

263
00:20:40,420 --> 00:20:44,420
which I think is gonna start a new season this month.

264
00:20:45,560 --> 00:20:46,000
Thank you.