Deepfakes

Manipulated media.

asisonline.org

What is a deepfake?

• Media (video, audio) that is manipulated by AI but seems authentic
• A mash-up of the words deep learning and fake

How do deepfakes work?

AI studies footage (pictures, videos, audio) of a person and then uses algorithms to generate new and convincing content.

What are the risks?

Bad actors can use deepfake tech to spread fake news and influence stock prices or elections, damage the credibility of influential people like politicians, CEOs or celebrities and scam people.

How to spot a deepfake?

1. Unnatural shadows/lighting
2. Disappearing ears
3. Blurred mouths/jawlines
4. Misaligned lips and audio
5. Changing skin tones

How long does it take to create a deepfake?

High-quality deepfakes require hours of footage consumption.

What is being done to combat deepfakes?

1. Facebook and Microsoft are collaborating with top US universities to create a deepfake database to help with detection and flagging.
2. Deeptrace (cybersecurity company) uses deep learning to detect if videos are manipulated or made via deepfake algorithms.
3. Twitter marks tweets that contain manipulated media and alerts users who want to engage with the tweet. Next to the deepfake, Twitter provides a link to a credible news article on a related topic. And, Twitter allows users to report tweets containing deepfakes.
4. Facebook hosted a Deepfake Detection Challenge for 2k+ participants who generated 35k+ models to help improve detection models. Facebook removes AI-generated media that was used to change a the content of a speech. And, Facebook labels as false any media that was edited to change the order or context of words.

How else can media outlets protect the public from deepfakes?

1. Work with lawmakers and enact regulations to disincentivize bad actors
2. Analyze virality and make it harder for deepfakes to spread

What is the catch?

The same algorithms used to detect flaws in deepfakes (Generative Adversarial Networks, or GANs), also helps to improve them and make them more believable. As the detection algorithm improves, so do the deepfakes.

Where do we see similar tech being used?

1. Faceswap — swap faces with a friend
2. Zao — swap faces with actors in iconic TV/movie clips
3. DataGrid — create a person from scratch (to be used for the fashion and apparel industry)

What is the future for deepfakes?

1. Researchers have developed systems that scrutinize videos for tells such as irregular blinking or unnatural shadows.
2. In 2020, the University of Buffalo published a paper explaining their successful detection technique to identify deepfakes by analyzing the reflection of light in eyes.
3. Use blockchain tech to verify the source before being able to post on social media platforms. Only content from trusted sources would be approved to help decrease the spread of potentially harmful deepfakes.

What now?

Be media literate and think critically:

1. Is this in line with what X person has said before?
2. How reliable is the source?
3. Has a major media outlet covered this?