W*E*L*C*O*M*E!
Fake News, Real HarmMultiple Perspectives on 'Disinformation'Dr. Saty Raghavachary, CS Dept, saty@usc.eduThis talk: bit.ly/AMS2019dis |
Disinformation - false information deliberately created and spread, with the sole intent to deceive - has the potential to rupture the fabric of society, because it strikes at the core notion of objective truth.
In this seminar, we will look at various forms of disinformation such as fake text, fake video, and fake audio, with plenty of examples of each, and will examine them from four perspectives:
* technical: use of deep learning, generative adversarial networks (GANs) and computer vision, for creating fake content (text, audio, video)
* cinematic: co-opting and extending Hollywood-style visual effects techniques, to create convincing video fakes
* journalistic: the threats and challenges that mainstream journalists and media face, given the ease of creation and viral propagation of fake media
* societal: educating the public about the dangers of mindlessly spreading fake news, and ways to critically examine and identify fake content
Attendees are expected to come away with a full understanding of fake media - what goes into their creation, why they are dangerous, and why we need to learn to counter its spread.
'Obviously' fake - but why?
'Fiction' is not "fact".
Fake news, for consumption at the supermarket checkout line.
CG and practical effects can 'bring back' dead people. Here is the Christian Dior commercial, 'starring' Grace Kelly and Marilyn Munroe.
For 99 cents...
Answer: intent.
The creators tell us they'd like to fool us - they ask us to willingly suspend disbelief, and we do.
All these ideas and tools for 'amusement purposes' can be easily "weaponized" into tools to spread DISINFORMATION.
Disinformation, not the same as misinformation - "false information which is intended to mislead, especially propaganda issued by a government organization to a rival power or the media."
Disinformation is created deliberately, ie. it is planned, and not an accident.
There is a vast array of deception techniques, all of which can be used to create disinformation.
It is possible to generate tweets in the style of actual people.
Given bland inputs, an entire story can be created, as can many more types of content.
Or, AI can create just story endings.
These faces do not exist.
If you want a set of 100,000 faces, here they are.
These interiors are... non-existent (and their descriptions are made up, too).
A 'fake video' was circulated, after the recent 737 Max 8 crash. The video by itself is not fake - the association is.
In other words, by simply associating content 'B' with an event 'A' ("repurposing" existing, unaltered content), via a simple tweet, the disinformation that INSTANTLY gets created is 'A caused B'.
It is so easy to discredit a grassroots movement, with the misuse of a single photograph.
After their 'defeat', ISIS seems to have switched to low-tech tactics to spread disinformation.
Again, effective low-tech tactics - anything goes - photos from another place/time, Photoshopped video grabs (from videogames!) etc.
'Deepfake' videos work by learning to transfer the 'style' (from multiple photographs) of person 'A', on to the 'content' (a single video) of person 'B' - a frame-by-frame face-swap operation that is automatic.
This is one of the earliest (SFW) examples.
ZAO, a Chinese deepfake video app, can graft a single photograph on to a video.
Disinformation creation can be done 'trivially' - a simple video edit (eg. time stretch filter) is enough to create confusion, forcing the victim to do immediate damage control.
"Modern", algorithmically-generated fakevideos (not doctored, existing videos), were first debuted on Reddit - a user named 'deepfakes' started posting doctored porn.
In Reddit, the 'deepfakes' subreddit contains a growing collection of deepfake videos.
Lyrebird analyzes ("learns") voice intonations associated with text, and uses it to create audio to match new text! Translation: text-to-speech, using ANYONE's voice and accent!!!
New video can be created by doctoring an existing clip, to match typed text.
SCARY.
Samsung researchers have come up with a technique to 'animate' a still face image/artwork/photo...
Hollywood's quest for photorealistic humans has a LONG history.
Some examples:
Paul Debevec @ USC has been a pioneer, in realistic illumination synthesis - this includes re-lighting-based CG animation.
Hao Li @ USC is another pioneer, in the area of point cloud processing, realistic CG hair generation (using input photographs), and lately, avatar generation.
AI, specifically, neural networks based learning, is the single biggest enabler of modern deepfake content.
A neural network (NN) "learns" (patterns latent in) its input data, by iterative 'weights' calculation.
Convolutional neural networks (CNNs) are a specific type of NN, where image-processing convolution filters are used in the network's layers.
Deep neural networks (DNNs) are simply, massively large NNs, with extreme complexity in their architecture - the complexity is their 'superpower'.
An NN is typically used to 'classify' ("detect") input data, ie. create a 'label' for it.
What if we instead create an NN, to GENERATE data instead? Further, what if we PAIR UP a generator and detector ("discriminator"), in an escalating game of robbers and cops? Result: GANs (Generative Adversarial Networks).
GANs are very powerful, and intriguing - there is an explosion of research in them.
It is possible to train an AI system, using LOTS of real-world text data, to calculate probabilities of word occurrences.
'pomo' [postmodern] is a system that spits out 'research articles' with perfect grammar.
This page has a clip at the top, that talks about the need for, and difficulties associated with, detection of fake videos.
This clip shows deepfake examples, and discusses some countering efforts.
This page summarizes a variety of aspects of deepfakes - technology, detection, dangers, etc.
One way - crowdsource the effort!
'Deepface Detection Challenge' is a Facebook initiative, to come up with state-of-the-art tools to identify fakes.
Problem: our desire to share consume, and share, indiscriminately.
Classically:
Today:
If the goal-keepers ("reputable" outlets, fact checkers...) are marginalized, or are simply overrun, disinformation will trump information.
Result: no "ground truth", so, no guarantee of stability in society! Creating discord, and polarization, is almost... trivial.
On top of technical solutions for deepface detection, we also need 'media literacy'.
Eg. in Italy, school children are taught about critical evaluation of media.
Facebook is making a commitment to combat disinformation.
You could test your own media savvy using the 'quiz' on this NYT page.
Compared to text, images or audio, videos have a more visceral effect on us; this combined with disinformation content, creation tools and distribution channels, makes facevideo a virulent combination (concept, and pic: Prof. Lurong Lei, Chongqing University):
Evaluating fake videos could be done, using these four dimensions (concept, and pic: Prof. Lurong Lei, Chongqing University):
Here is a 'capability pyramid' (concept, and pic: Prof. Lurong Lei, Chongqing University) that can be used to become knowledgeable on detecting and dealing with fakevideos:
Here is a doc that summarizes the development of deepfakes, and includes a history timeline.