Op-ed: Deepfakes — Don’t believe anything you see on the internet

Source: Artificial Intelligence on Medium

Op-ed: Deepfakes — Don’t believe anything you see on the internet

By Ujjwal Singhania, MEng ’20 (EECS)

One of the images below is real and the other has been manipulated by a computer (known as a deepfake image). Take a few seconds and try to figure out which one is which.

Image 1 and 2 — One of these images is a deepfake, can you tell which one?

If you thought the image on the left was real, you have been successfully fooled by a deepfake.

The proliferation of artificial intelligence has led to the rise of deepfakes. Deepfakes are altered images and videos that replace people in existing media with someone else using neural networks. Deepfakes can also involve editing the audio of a video to make an actor say things they never said. Recently, they are being used to manipulate voters and sway elections, to create revenge porn and to discredit influential figures on the internet.

The proliferation of artificial intelligence has led to the rise of deepfakes.

How are they created?

The two main ways to create deepfakes are through the use of artificial intelligence (more specifically generative adversarial networks) or the use of image/audio editing software. Due to widely available cheap compute, methods using AI have been able to create hyper-realistic deepfakes that are very difficult to detect. Software created deepfakes, on the other hand, take less time to create but are more easily detectable compared to AI-created deepfakes.

Several techniques are used to create deepfakes. The most common way is by doing a ‘FaceSwap’. This technique involves detecting a face in the original media and replacing it with another face using computational magic. Due to the copy and paste nature of this technique, it is difficult to generate deepfakes with different skin tones and genders. An advanced technique to create deepfakes is ‘Face2Face’. Face2Face enables the “real-time face capture and reenactment of videos,”2 which results in the creation of deepfakes that are very difficult to detect. The following video demonstrates Face2Face.

Is it easy to create a deepfake?

Due to the easy availability of cheap compute through Google Cloud and Amazon Web Services, it has become very easy to create deepfakes. Many popular deepfake generation tools and algorithms are also available to download for free on websites such as GitHub and GitLab. Due to this ubiquity, anyone with basic knowledge about deepfakes and a computer with an internet connection can create a photo-realistic deepfake in hours.

How to detect deepfakes on the Internet?

Companies such as Facebook, Google and Microsoft, and academic institutions such as UC Berkeley, and MIT are heavily investing to fund the creation of deepfake detectors that can accurately detect and flag deepfakes on the internet. A fundamental obstacle with deepfake detection is that due to the large variety of techniques that can be used to generate deepfakes, it is difficult to create a robust classifier that generalizes well. Deepfakes generated by new and novel methods can slip through the classifier’s detection process and be exposed to a vast online audience.

Our capstone team at UC Berkeley is working on creating a generalized classifier that is capable of detecting deepfakes created via different methods, including those that the classifier hasn’t seen before. If successful, our robust classifier could be used to filter deepfakes from social media, and search results, and prevent the spread of misinformation that degrades trust and privacy.

What are some positives applications of deepfakes?

Like all technology, deepfakes have both positive and malicious applications. Constructive applications of deepfakes include the ability to dub a movie into different languages, as well as de-aging technology that was most recently used by the Netflix original, The Irishman. Deepfakes can also be useful in the early application of futuristic communication protocols such as holograms.

The question on your mind now is, probably, what can you, as a consumer, do to protect yourself from deepfakes? The absence of a robust deepfake detector makes it difficult for companies and consumers alike to be shielded from deepfakes. However, companies such as Facebook have very lax standards concerning the spread of misinformation using deepfakes. Facebook, in particular, refuses to delete deepfakes from its platform (a notable example being the viral Nancy Pelosi and Mark Zuckerberg deepfake)3. We can nudge these companies to have stricter guidelines and regulate content on their websites with a heavier emphasis on privacy. As an internet user, one routinely hears, “don’t believe anything you see on social media.” In the era of deepfakes, my key advice would be — don’t believe anything you see on the internet.