What do deep fakes mean for our First Amendment protections? You were asked to consider this question while reviewing the powerpoint on the technicalities of deep-faking.

Deepfakes: Trick or Treat?

By: Kietzmann, J., Lee, L.W., McCarthy, I.P. and Kietzmann, T.C.

In the journal: Business Horizons.

The authors

Jan H. Kietzmann Ian P. McCarthyLinda W. Lee Tim C. Kietzmann

Deepfakes: Trick or Treat?

The paper

Click here to access the paper

What are deepfakes?

“Deepfakes leverage powerful techniques from machine learning and artificial intelligence to manipulate or generate visual and audio content with a high potential to deceive” (Kietzmann et al. 2020).

Example: Rowan Atkinson (Mr Bean) unexpectedly stars in a perfume commercial (original recorded with Charlize Theron).

View the original advert here: https://youtu.be/VqSl5mSJXJs

View the deepfake here: https://youtu.be/tDAToEnJEY8

Deepfakes: Trick or Treat?

What are deepfakes?

The phenomenon gained its name from a user of the platform Reddit, who went by the name “deepfakes” (deep learning + fakes).

This person shared the first deepfakes by placing unknowing celebrities into adult video clips. This triggered widespread interest in the Reddit community and led to an explosion of fake content.

The first targets of deepfakes were famous people, including actors (e.g., Emma Watson and Scarlett Johansson), singers (e.g., Katy Perry) and politicians (e.g., President Obama)

Deepfakes: Trick or Treat?

Deepfakes matter because:

Believability: If we see and hear something with our own eyes and ears, we believe it to exist or to be true, even if it is unlikely.

The brain’s visual system can be targeted for misperception, in the same way optical illusions and bistable figures trick our brains.

Deepfakes: Trick or Treat?

Deepfakes matter because:

Accessibility: the technology of today and tomorrow, will allow all of us to create fakes that appear real, without a significant investment in training, data collection, hardware and software.

Zao, the popular Chinese app for mobile devices lets users place their faces into scenes from movies and TV shows, for free.

Deepfakes: Trick or Treat?

How do deepfakes work?

Consider below deepfake featuring Jim Carrey and Alison Brie:

The original Alison Brie video: https://www.youtube.com/watch?v=QBmYDzLhWoY

The deepfake with Jim Carrey: https://www.youtube.com/watch?v=b5AWhh6MYCg Deepfakes: Trick or Treat?

How do deepfakes work?

Many deepfakes are created by a three-step procedure:

Deepfakes: Trick or Treat?

How do deepfakes work?

Step 1: The image region showing Brie’s face is extracted from an original frame of the video. This image is then used as input to a deep neural network (DNN), a technique from the domain of machine learning and artificial intelligence..

Step 2: The DNN automatically generates a matching image showing Carrey instead Brie.

Step 3: This generated face is inserted into the original reference image to create the deepfake.

Deepfakes: Trick or Treat?

How do deepfakes work?

The main technology for creating deepfakes is deep learning, a machine learning method used to train deep neural networks (DNNs).

DNNs consist of a large set of interconnected artificial neurons, referred to as units.

Much like neurons in the brain, while each unit itself performs a rather simple computation, all units together can perform complex nonlinear operations.

In case of deepfakes, this is mapping from an image of one person to another.

Deepfakes: Trick or Treat?

How do deepfakes work?

Deepfakes are commonly created using a specific deep network architecture known as autoencoder.

Autoencoders are trained to recognize key characteristics of an input image to subsequently recreate it as their output. In this process, the network performs heavy data compression.

Autoencoders consist of three subparts:

– an encoder (recognizing key features of an input face) – a latent space (representing the face as a compressed version) – a decoder (reconstructing the input image with all detail)

Deepfakes: Trick or Treat?

How do deepfakes work?

Autoencoder: a DNN architecture commonly used for generating deepfakes.

Deepfakes: Trick or Treat?

How do deepfakes work?

Encoder: Much like an artist drawing an image, the encoder compresses an image, from originally tens of thousands of pixels into a few hundred (typically around 300) measurements.

These measurements can relate to particular facial characteristics e.g., eyes are open or closed, the head pose, the emotional expression, skin colour, etc.

Deepfakes: Trick or Treat?

How do deepfakes work?

Latent space: represents different facial aspects of the person on which it is trained.

It is often compared to information bottlenecks so that the network can learn more general facial characteristics rather than memorizing all input examples of specific people.

The compression achieved by the encoding of an input image into the latent space can be 0.1% of the memory needed to store the original input image.

Deepfakes: Trick or Treat?

How do deepfakes work?

Decoder: decompresses the information in the latent space to reconstruct an image as perfectly as possible.

The performance of the whole autoencoder network is measured by how much the input and generated (output) images resemble each other. This task is made difficult because of the heavy data compression performed by the encoder.

Deepfakes: Trick or Treat?

The deepfake trick

Two separate autoencoders trained each on a different person will be very different and cannot be integrated.

The trick for creating deepfakes lies in sharing the encoder across two networks such that they remain compatible.

This way, the image of one person can be used to compute a compressed latent space representation, from where the decoder of another person is used to create the fake.

Deepfakes: Trick or Treat?

The deepfake trick

Deepfakes: Trick or Treat?

Using the same encoder and hence latent space representation for images of two separate people is key to understanding deepfakes.

If two autoencoders were trained separately, the latent spaces would not be aligned (Brie, and Carrey latent space below). Encoder sharing will result in an aligned latent space (grey dots). The autoencoders can then be used to match from one to another person.

The deepfake trick

Deepfakes: Trick or Treat?

A shared encoder is key to creating novel facial images of a target person that exhibit the same emotional expression, head posture, etc. as the original facial characteristics. This new image can then be doctored back into the original image to create a fake scene.

A typology of deepfakes and their business applications

Type Description Current example Business application

Photo deepfakes

Face and body-swapping Making changes to a face, replacing or blending the face (or body) with someone else’s face (or body

FaceApp’s aging filter alters your photo to show how you might look decades from now (Kaushal, 2019).

Consumers can virtually try on cosmetics, eye glasses, hairstyles or clothes.

Audio deepfakes

Voice-swapping Changing a voice or imitating someone else’s voice.

Text to Speech Changing audio in a recording by typing in new text

Fraudsters used AI to mimic a CEO’s voice and then tricked a manager into transferring $243,000 (Supasorn Suwajanakorn, 2017).

Users could make controversial Dr. Jordan B. Peterson a famous professor of psychology and author say anything they wanted, until his threat of legal action shut the site NotJordanPeterson down (Cole, 2019).

The voice of an audio book narration can sound younger, older, male, or female and with different dialects or accents to take on different characters.

Misspoken words or a script change in a voiceover can be replaced without making a new recording.

Deepfakes: Trick or Treat?

A typology of deepfakes and their business applications

Type Description Current example Business application

Video deepfakes

Face-swapping Replacing the face of someone in a video with the face of someone else.

Face-morphing A face changes into another face through a seamless transition

Full body puppetry Transposing the movement from one person’s body to that of another.

Jim Carrey’s face replaces Alison Brie’s in “Late Night with Seth Meyers” interview.

Former “Saturday Night Live” star Bill Hader imperceptibly morphs in and out of Arnold Schwarzenegger in the talk show Conan.

“Everybody dance now” shows how anyone can look like a professional dancer.

Face-swapped video can be used to put the leading actor’s face onto the body of a stunt double for more realistic-looking action shots in movies.

Video game players can insert their faces onto that of their favorite characters.

Business leaders and athletes can hide physical ailments during a video presentation.

Deepfakes: Trick or Treat?

The R.E.A.L. framework for dealing with the darkside of deepfakes

Deepfakes: Trick or Treat?

The R.E.A.L. framework

Record: Deepfakes often seek to falsely portray somebody doing or saying something and being somewhere, the exposure of such fakes would require evidence (or an “alibi”) to the contrary.

Technology already exists to track and “life-log” a person’s life in terms of location, communications and activities.

Such life-log data could then be encrypted, stored and used to help identify and expose the posting of dark deepfakes.

Deepfakes: Trick or Treat?

The R.E.A.L. framework

Expose: Technological innovations are being developed to detect and classify deepfakes by identifying issues with image resolution, scaling, rotation and splicing.

The U.S.’s Defense Advanced Research Projects Agency (DARPA) has a Media Forensics program.

Reuters developed a free online tutorial to help us identify manipulated media such as deepfakes.

The Deepfake Detection Challenge invites people around the world to build innovative new technologies that can help detect deepfakes and manipulated media. Deepfakes: Trick or Treat?

The R.E.A.L. framework

Advocate: At the moment there is little legal consequence for producing, hosting and sharing deepfakes. This is changing.

In China, as 1st January 2020, it is a criminal offense to publish deepfakes or fake news without disclosure.

Also victims have legal recourse in instances of:

● defamation, malice, breaches of privacy or emotional distress from a deepfake, and

● cases of copyright infringements, impersonation and fraud involving deepfakes.

Deepfakes: Trick or Treat?

The R.E.A.L. framework

Leverage: One way to counter deepfakes is for individuals and organizations to strengthen trust in their content and presence.

Individuals and organizations with strong and respected brands will be better positioned to weather deepfake assaults, as their stakeholders will defend their brand.

When brands built on strong ethics are portrayed in an unfavorable light in deepfakes, the hope is that stakeholders will not simply believe their eyes and ears, but be more critical and think for themselves.

Deepfakes: Trick or Treat?

Takeaways

Deepfakes can be used in positive and negative ways to manipulate content for media, entertainment, marketing and education.

Increasingly our lives are being captured via social media and this content can be used to train DNNs, with or without our permission.

Deepfakes are not magic, but are produced using techniques from AI that can generate fake content that is highly believable.

Deepfakes: Trick or Treat?

The DOI (Digital Object Identifier) for the paper on which these slides are based:

https://doi.org/10.1016/j.bushor.2019.11.006