-
Notifications
You must be signed in to change notification settings - Fork 1
Expand file tree
/
Copy pathindex.html
More file actions
414 lines (367 loc) · 25.3 KB
/
index.html
File metadata and controls
414 lines (367 loc) · 25.3 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1, shrink-to-fit=no">
<meta name="description" content="">
<meta name="author" content="">
<link rel="shortcut icon" href="favicon.ico" type="image/x-icon">
<link rel="icon" href="favicon.ico" type="image/x-icon">
<title>equalAIs</title>
<!-- Bootstrap core CSS -->
<link href="vendor/bootstrap/css/bootstrap.min.css" rel="stylesheet">
<!-- Custom fonts for this template -->
<link href="vendor/font-awesome/css/font-awesome.min.css" rel="stylesheet" type="text/css">
<link href="https://fonts.googleapis.com/css?family=Lora:400,700,400italic,700italic" rel="stylesheet" type="text/css">
<link href='https://fonts.googleapis.com/css?family=Cabin:700' rel='stylesheet' type='text/css'>
<!-- Custom styles for this template -->
<link href="css/equalais.css" rel="stylesheet">
</head>
<body id="page-top">
<!-- Navigation -->
<nav class="navbar navbar-expand-lg navbar-light fixed-top" id="mainNav">
<div class="container">
<img src="img/Logos/MiniLogo.png" class="img-fluid mr-2" alt="equalAIs Logo">
<a class="navbar-brand js-scroll-trigger" href="#tool">REMOVE THE BOX</a>
<button class="navbar-toggler navbar-toggler-right" type="button" data-toggle="collapse" data-target="#navbarResponsive" aria-controls="navbarResponsive" aria-expanded="false" aria-label="Toggle navigation">
Menu
<i class="fa fa-bars"></i>
</button>
<div class="collapse navbar-collapse" id="navbarResponsive">
<ul class="navbar-nav ml-auto">
<li class="nav-item">
<a class="nav-link js-scroll-trigger" href="#page-top">Home</a>
</li>
<li class="nav-item">
<a class="nav-link js-scroll-trigger" href="#about">About</a>
</li>
<li class="nav-item">
<a class="nav-link js-scroll-trigger" href="#faq">FAQ</a>
</li>
<li class="nav-item">
<a class="nav-link js-scroll-trigger" href="#team">Team</a>
</li>
<li class="nav-item">
<a class="nav-link js-scroll-trigger" href="#contact">Contact</a>
</li>
</ul>
</div>
</div>
</nav>
<!-- Intro Header -->
<header class="masthead">
<div class="intro-body">
<div class="container">
<div class="row">
<div class="col-lg-8 mx-auto">
<p><img src="img/Logos/Logo.png" width="25%"class="img-fluid" alt="equalAIs Logo"></p>
<h1 class="brand-heading">equalAIs</h1>
<h3><p class="intro-text">Empowering humans by subverting machines</p></h3>
<a href="#about" class="btn btn-circle js-scroll-trigger">
<i class="fa fa-angle-double-down animated"></i>
</a>
</div>
</div>
</div>
</div>
</header>
<!-- About Section -->
<section id="about" class="content-section-eqlesspadding text-center">
<div class="container text-justify">
<h2 class="text-center">About</h2>
<h3 class="text-center">The Problem</h3>
<p>What are the policy and societal implications of the unprecedented capability for automated, real time identification and tracking of individuals?</p>
<p>Private and public surveillance cameras combined with cheap, pervasive, automated facial recognition and classification technology have made it possible to know, in an automated and scalable way, who a person associates with, where they go and when, what their habits and likes and dislikes are and what their mood and health status is. For the first time in the U.S., the government is compiling a national biometric database far beyond known or even suspected criminals. 80% of the facial database entries are unrelated to arrests, from sources including driver’s licenses and passports. The largest private international tech companies have even bigger facial databases and more advanced facial identification algorithms.</p>
<p>Ubiquitous automated surveillance by the government is likely to have a chilling effect on protected free speech, association and religious activities and weaken the constitutional protections for citizens accused of crimes.</p>
<p>Access to the same data by private actors leads to a host of privacy and policy questions. Foreseeable bad outcomes for individuals include identity theft, stalking, psychological manipulation and reputational and economic damage.</p>
<p class="text-left"><a href="https://www.priv.gc.ca/en/opc-actions-and-decisions/research/explore-privacy-research/2013/fr_201303/">Automated Facial Recognition in the Public and Private Sectors</a></p>
<p class="text-left"><a href="https://www.ftc.gov/system/files/documents/public_events/1223263/privacycon_worldofcameras_das2_0.pdf">Assisting Users in a World Full of Cameras: A Privacy-aware Infrastructure for Computer Vision Applications</a></p>
</div>
</section>
<!-- About Section -->
<section id="project" class="content-section-lesspadding text-center">
<div class="container text-justify">
<h3 class="text-center">The Solution</h3>
<p>We believe there should be more public discourse around what
choices we want to make as individuals and as a society regarding how our data and
images and facial recognition will be used and more tools to make those choices and
to express or enforce them. To that end, our work has multiple components, from a
FOIA request for information about the federal government’s use of facial recognition
to an open sourced technical prototype for changing photographs so that facial
recognition systems can’t see faces in them.</p>
<p>1. Policy Investigation
Examine the policy and societal implications of the unprecedented capability for automated,
real time identification and tracking of individuals resulting from the increasing ubiquity of
private and public surveillance cameras combined with cheap, pervasive, automated facial
recognition and classification technology.</br>
● FOIA request</p>
<p>2. Technical Adversarial Attack Prototype
Show technical feasibility of facial recognition adversarial attacks through creation and
improvement of a technical prototype.</br>
● Prototype based on Cleverhans and Goodfellow, et. al.</p>
<p>3. Communication and Raising Awareness
Stimulate public awareness and discourse about those policy implications. Presentations at:</br>
● AMLC TomTomFest</br>
● The American Academy of Arts and Science</br>
● Harvard Berkman Klein Center</p>
<p>4. Build Open-Source Community
Create an open source repository of code, papers and other resources to facilitate further
discussion, research and experimentation on facial recognition, facial recognition adversarial
attacks and related policy questions by a broad set of stakeholders (e.g. production engineers
as well as academics.)</br>
● <a href="https://github.com/equalais/assembly_equalais">Code and resources</a> </p>
<p>5. Industry Partnerships, Enforcement & Regulation Investigate possible policy alternatives to the status quo, such as an opt in or an opt out system to
avoid face tagging on media platforms. Consider industry partnerships and voluntary standards or regulatory or other enforcement mechanisms.</br>
● Multi layered filter: steganographic embedding, human and machine readable “do not read” overlay and adversarial perturbation</p>
<h3 class="text-left"><small>equalAIs System Diagram</small></h3>
<p><img src="img/DiagramRounded.png" class="img-fluid" alt="equalAIs System Diagram"></p>
</div>
</section>
<!-- Tool Section -->
<section id="tool" class="tool-section content-section-eqlesspadding text-center">
<div class="container">
<div class="row">
<div class="col-lg-8 mx-auto">
<h2>THE TOOL</h2>
<p>1. Browse for a picture you would like to EqualAIs (prevent face detection)</br>
<label class="btn btn-default btn-lg" for="file">BROWSE</label>
<image id='loading' src="img/loader.gif" style="display: none;" alt="loader">
</p>
<form id="uploadimage" action="" method="post" enctype="multipart/form-data">
<input type="file" name="image" class="btn btn-default btn-lg" style="display: none;" id="file" required />
<input type="submit" value="Upload" class="submit" style="display: none;"/>
</form>
<p>2. You will get a pop up asking for confirmation to download the new result
picture (on some browsers download will commence automatically).
</p>
<p>3. Use it on social media and spread the word!</p>
<p>You can find a link to our repository here:</p>
<ul class="list-inline banner-social-buttons">
<li class="list-inline-item">
<a href="https://github.com/equalais/assembly_equalais" class="btn btn-default btn-lg">
<i class="fa fa-github fa-fw"></i>
<span class="network-name">Github</span>
</a>
</li>
</ul>
</div>
</div>
</div>
</section>
<section class="content-section-lesspadding text-center" id="faq">
<div class="container">
<h2>FAQ</h2>
<h3 class="text-left"><small>Q. What do you mean by adversarial attack?</small></h3>
<p class="text-left text-justify">A. An adversarial attack is a perturbation deliberately applied to inputs in order to change classifier predictions while leaving the content easily recognizable to humans. Deep learning classifiers are known to be susceptible to adversarial attacks, with examples in previous work including fooling models into saying a stop sign isn’t there, a cat is guacamole or a toy turtle is a rifle. In our case, the classifiers are neural network models that have been trained to find faces in images. The adversarial attack is a perturbation of the images such that the faces in the perturbed images are still easily recognizable to humans but the model can no longer see them and reports with high confidence that the perturbed images contain no faces.</br>
<a href="https://arxiv.org/abs/1312.6199">Intriguing properties of neural networks</a>
</p>
<h3 class="text-left"><small>Q. How is an adversarial attack created?</small></h3>
<p class="text-left text-justify">A. Most adversarial examples are made using the gradient of the neural network model. At each pixel, the algorithm tests what direction of change in pixel value makes the probability of the class we're trying to avoid, in our case "face", decrease the most (another set of adversarial attack approaches is designed to target a certain class, so in that case the direction chosen is the one that most increases the probability of the target class.) The algorithm then makes a slight change in that direction. Because it’s a high dimensional space and neural networks are designed to act very linear to make them easy to optimize, the many small input changes add up to a large change in the output, reliably creating examples that are misclassified by a wide variety of models.</br>
<a href="https://arxiv.org/abs/1412.6572">Explaining and Harnessing Adversarial Examples</a>
</p>
<h3 class="text-left"><small>Q. Why an adversarial attack?</small></h3>
<p class="text-left text-justify">A. Adversarial attacks work against the most common kinds of image classifiers.</br>
An adversarial attack, without deliberate remedial action by the owner of the classifier, makes facial identification and recognition technically impossible, so it has the advantage of not relying on active cooperation by any other parties.</br>
Even if it works only partially or briefly, its existence supports a conversation about whether we want to and should be able to have choices about how our society is using biometric data, facial recognition and surveillance.
</p>
<h3 class="text-left"><small>Q. Is this a cat and mouse game? Why is it hard to defend (against adversarial examples)?</small></h3>
<p class="text-left text-justify">A. Given a particular type of adversarial example, a defender can add examples of that type to their training data and improve their model’s performance, but this adversarial training only improves performance against that particular type of adversarial example. The model will be vulnerable to the next type of adversarial example the attacker comes up with and generally has slightly worse performance than the original model on clean data. No one has designed a strong, general defense algorithm that can resist a wide variety of attacks and no one knows whether it’s theoretically possible, or whether fundamentally the defender or the attacker has the advantage if they both knew their optimal strategies.</br>
One reason it’s hard to defend against adversarial examples is that the space of all possible images is enormous and any sample of real data, even the entirety of all photographs ever taken, is an extremely small fraction. There are many, many ways to change a photograph slightly, leaving it easily recognizable to humans, and many of those shift the classifier’s results away from the ground truth label and towards something else. How many, we don’t really know.</br>
Adversarial examples are solutions to a complicated optimization problem where we don’t yet have the theoretical tools to describe that problem’s solution space. So while we can generate and defend against known types, in general, we don’t know how many types there are, what properties they have or how to generate or defend against them.</br>
<a href="http://www.cleverhans.io/security/privacy/ml/2017/02/15/why-attacking-machine-learning-is-easier-than-defending-it.html">Why attacking machine learning is easier than defending it</a>
</p>
<h3 class="text-left"><small>Q. What are the limitations of an adversarial attack?</small></h3>
<p class="text-left">A. An adversarial attack can provide at best partial protection against all of the uses others might make of an individual’s image. It is limited in which images it will be applied to and how long the attack will be effective. For type of image, it is only applied to images the individual places into a system. It has no effect on images friends, acquaintances, employers, strangers, store owners, property managers, government agencies or anyone else places into the system and photographs of our faces are routinely taken and posted publicly without our knowledge or consent. For length of effectiveness, a specific kind of adversarial attack will only work as long as the owner of the classification model is ignorant of or indifferent to it. If it’s discovered, the model can be retrained to classify that kind of adversarial example correctly. A third limitation is the degradation of the image quality, lowering its utility for the individual who applied the adversarial filter. Some adversarial attacks are so subtle that the change isn’t noticeable to a human observer, but others make very noticeable changes even though the subject is still recognizable.
</p>
<h3 class="text-left"><small>Q. What are the results of our adversarial attacks?</small></h3>
<p><img src="img/AA_Results.png" class="img-fluid" alt="Adversarial Attack Results"></p>
<h3 class="text-left"><small>For more information see:</small></h3>
<p class="text-left"><a href="http://www.cleverhans.io">http://www.cleverhans.io</a></p>
<h3 class="text-left"><small>References:</small></h3>
<p class="text-left"><a href="http://thehill.com/blogs/pundits-blog/technology/341906-opinion-facial-recognition-surveillance-is-here-but-privacy">Facial recognition surveillance is here — but privacy protections are not</a></p>
<p class="text-left"><a href="https://www.washingtonpost.com/news/the-switch/wp/2015/06/11/facial-recognition-technology-is-everywhere-it-may-not-be-legal/">Facial recognition technology is everywhere. It may not be legal.</a></p>
<p class="text-left"><a href="https://www.aclu.org/issues/privacy-technology/surveillance-technologies/face-recognition-technology">ACLU: Surveillance technologies —— face recognition technology</a></p>
<p class="text-left"><a href="https://www.theguardian.com/technology/2017/mar/27/us-facial-recognition-database-fbi-drivers-licenses-passports">Facial recognition database used by FBI is out of control</a></p>
<p class="text-left"><a href="https://www.nytimes.com/2018/03/26/technology/ftc-facebook-investigation-cambridge-analytica.html">Demands Grow for Facebook to Explain Its Privacy Policies</a></p>
<p class="text-left"><a href="https://techcrunch.com/2018/04/16/judge-says-class-action-suit-against-facebook-over-facial-recognition-can-go-forward/">Judge says class action suit against Facebook over facial recognition can go forward</a></p>
</div>
</section>
<!-- Team Section -->
<section id="team" class="content-section-lesspadding text-center">
<div class="container">
<div class="row">
<div class="col-lg-8 mx-auto">
<h2>THE TEAM</h2>
</div>
</div>
<!-- Team Members Row -->
<div class="row">
<div class="col-md-4 text-center mb-4">
<img id="dp" class="rounded-circle img-fluid mx-auto mb-4" src="img/Faces/pedraza_daniel_s_p.png" alt="pedraza_daniel">
<h3>Daniel Pedraza</br>
<small>Project Manager</small></br>
<a href="https://www.linkedin.com/in/sumdp/" class="btn-lg">
<i class="fa fa-linkedin fa-fw"></i>
</a>
<a href="https://github.com/sumdp" class="btn-lg">
<i class="fa fa-github fa-fw"></i>
</a>
<a href="https://twitter.com/sumdp" class="btn-lg">
<i class="fa fa-twitter fa-fw"></i>
</a>
<a href="mailto:me@dpedraza.com" class="btn-lg">
<i class="fa fa-envelope fa-fw"></i>
</a>
</h3>
</div>
<div class="col-md-4 text-center mb-4">
<img id="da" class="rounded-circle img-fluid mx-auto mb-4" src="img/Faces/adjodah_dhaval_s_p.png" alt="adjodah_dhaval">
<h3>Dhaval Adjodah</br>
<small>Machine Learning Engineer</small></br>
<a href="https://www.linkedin.com/in/rand0m/" class="btn-lg">
<i class="fa fa-linkedin fa-fw"></i>
</a>
<a href="https://github.com/d-val" class="btn-lg">
<i class="fa fa-github fa-fw"></i>
</a>
<a href="http://twitter.com/_dval_" class="btn-lg">
<i class="fa fa-twitter fa-fw"></i>
</a>
<a href="mailto:dhaval@mit.edu" class="btn-lg">
<i class="fa fa-envelope fa-fw"></i>
</a>
</h3>
</div>
<div class="col-md-4 text-center mb-4">
<img id="gg" class="rounded-circle img-fluid mx-auto mb-4" src="img/Faces/greene_gretchen_s_p.png" alt="greene_gretchen">
<h3>Gretchen Greene</br>
<small>Machine Learning Engineer</small></br>
<a href="https://www.linkedin.com/in/kgretchengreene/" class="btn-lg">
<i class="fa fa-linkedin fa-fw"></i>
</a>
<a href="mailto:kgretchengreene@gmail.com" class="btn-lg">
<i class="fa fa-envelope fa-fw"></i>
</a>
</h3>
</div>
<div class="col-md-4 text-center mb-4">
<img id="jj" class="rounded-circle img-fluid mx-auto mb-4" src="img/Faces/joseph_josh_s_p.png" alt="joseph_josh">
<h3>Josh Joseph</br>
<small>Machine Learning Engineer</small></br>
<a href="https://www.linkedin.com/in/jmjoseph/" class="btn-lg">
<i class="fa fa-linkedin fa-fw"></i>
</a>
<a href="mailto:jmjoseph@alum.mit.edu" class="btn-lg">
<i class="fa fa-envelope fa-fw"></i>
</a>
</h3>
</div>
<div class="col-md-4 text-center mb-4">
<img id="tm" class="rounded-circle img-fluid mx-auto mb-4" src="img/Faces/miano_thom_s_p.png" alt="miano_thom">
<h3>Thom Miano</br>
<small>Machine Learning Engineer</small></br>
<a href="https://www.linkedin.com/in/tnmiano/" class="btn-lg">
<i class="fa fa-linkedin fa-fw"></i>
</a>
<a href="https://github.com/thommiano" class="btn-lg">
<i class="fa fa-github fa-fw"></i>
</a>
<a href="http://twitter.com/SocraticDatum" class="btn-lg">
<i class="fa fa-twitter fa-fw"></i>
</a>
<a href="mailto:thom.miano@gmail.com" class="btn-lg">
<i class="fa fa-envelope fa-fw"></i>
</a>
</h3>
</div>
<div class="col-md-4 text-center mb-4">
<img id="f" class="rounded-circle img-fluid mx-auto mb-4" src="img/Faces/Francisco_s_p.png" alt="Francisco">
<h3>Francisco</br>
<small>Machine Learning Engineer</small></br>
</h3>
</div>
</div>
<div class="row">
<div class="col-lg-8 mx-auto">
<a href="team.html" class="btn btn-default btn-lg">MORE</a>
</div>
</div>
</div>
</section>
<!-- Contact Section -->
<section id="contact" class="contact-section content-section text-center">
<div class="container">
<div class="row">
<div class="col-lg-8 mx-auto">
<h2>Contact the team:</h2>
<p>Feel free to leave us a comment or to give some feedback about this project!</p>
<ul class="list-inline banner-social-buttons">
<li class="list-inline-item">
<a href="https://twitter.com/Equal41s" class="btn btn-default btn-lg">
<i class="fa fa-twitter fa-fw"></i>
<span class="network-name">Twitter</span>
</a>
</li>
<li class="list-inline-item">
<a href="https://github.com/equalais" class="btn btn-default btn-lg">
<i class="fa fa-github fa-fw"></i>
<span class="network-name">Github</span>
</a>
</li>
<li class="list-inline-item">
<a href="mailto:equalais@media.mit.edu" class="btn btn-default btn-lg">
<i class="fa fa-envelope fa-fw"></i>
<span class="network-name">EMAIL</span>
</a>
</li>
</ul>
</div>
</div>
</div>
</section>
<!-- Attributions -->
<section class="content-section-fewpadding text-center" id="attributions">
<div class="container">
<h2>Acknowledgements</h2>
<p class="mb-5 text-justify">We are thankful to the Berkman Klein Center and MIT Media Lab for having
had the opportunity to develop the equalAIs project under the 2018 Assembly
program. In particular, we are thankful to the Assembly administrators,
board of advisors, Harvard Cyber Law Clinic (Kendra Albert), and other
Assembly cohort members who have provided invaluable guidance and support
along the way. We also appreciate the support and mentorship of Kade Crockford,
James Atwood, and Jessy Lin.</p>
<div class="row mt-5">
<div class="col-md-4 mx-auto mb-5">
<a href="https://cyber.harvard.edu/">
<img src="img/Logos/BKC.png" class="img-fluid" alt="BKC">
</a>
</div>
<div class="col-md-4 mx-auto mb-5">
<a href="https://bkmla.org/">
<img src="img/Logos/assemblylogo_white.png" class="img-fluid" alt="Assembly">
</a>
</div>
<div class="col-md-4 mx-auto">
<a href="https://www.media.mit.edu/">
<img src="img/Logos/Mit_medialab_logo.png" class="img-fluid" alt="Media Lab">
</a>
</div>
</div>
</div>
</section>
<!-- Footer -->
<footer>
<div class="container text-center">
<p>Copyright © equalAIs 2018</p>
</div>
</footer>
<!-- Bootstrap core JavaScript -->
<script src="vendor/jquery/jquery.min.js"></script>
<script src="vendor/bootstrap/js/bootstrap.bundle.min.js"></script>
<!-- Plugin JavaScript -->
<script src="vendor/jquery-easing/jquery.easing.min.js"></script>
<!-- Custom scripts for this template -->
<script src="js/equalais.js"></script>
<script src="js/tool.js"></script>
</body>
</html>