AI Deepfakes

- Overview

AI deepfakes are realistic videos, images, audio, or text that are created using artificial intelligence (AI) to depict events that never happened. The term "deepfake" combines the concepts of deep learning (DL) and something fake. 

Deepfakes are created by using machine learning (ML) algorithms to stitch together hoaxed images and sounds. The process typically involves feeding hundreds or thousands of images into an artificial neural network (ANN) to train it to recognize and reconstruct patterns, usually faces. 

Deepfakes can have serious consequences, including defamation, privacy violations, and the fabrication of criminal evidence.

AI effectively learns what a source face looks like at different angles in order to transpose the face onto a target, usually an actor, as if it were a mask. Huge advances came through the application of generative adversarial networks (GANS) to pit two AI algorithms against each other, one creating the fakes and the other grading its efforts, teaching the synthesis engine to make better forgeries. 

Hollywood has transposed real or fictional faces onto other actors, for example, bringing Peter Cushing back to life in 2016’s Rogue One: A Star Wars Story, but the technique used complex, expensive pipelines and face-mounted cameras. 

- Applications of AI Deepfake Technology

Deepfakes can be used for many purposes, including: 

• Misleading the public: Deepfakes can be used to spread false information or propaganda, such as by showing a celebrity or world leader saying something they never said.
• Entertainment: Deepfakes can be used for fun, such as in art installations that allow users to take a selfie with Salvador Dalí.
• Breaking down linguistic barriers: Deepfakes can be used to translate messages into multiple languages.
• Helping people with disabilities: Deepfakes can be used to allow people with disabilities to communicate, such as by using deepfake technology to allow Val Kilmer to "speak" after losing his voice to throat cancer.

- Deepfake Voice Technology

Deepfake voice technology allows people to spoof the voices of other people - often politicians, celebrities or CEOs - using AI. As the world of AI and deepfake technology grows more complex, the risk that deepfakes pose to firms and individuals grows increasingly potent. 

This growing sophistication of the latest software and algorithms has allowed malicious hackers, scammers and cyber criminals who work tirelessly behind the scenes to stay one step ahead of the authorities, making the threat of attacks increasingly difficult to both prepare for and defend against.  

The audio deepfake scam is, without a doubt, one of the more bizarre applications of deepfake technology. However, as we’ve seen, it’s one which can clearly be applied successfully ­– so successfully and convincingly, in fact, that the CEO who fell victim to the cyberattack stated on the record that he recognized his boss’s voice by its ‘slight German accent’ and ‘melodic lilt.’  Furthermore, by all accounts, the cybercriminals’ tech is becoming more difficult to detect by the month. 

Sophisticated technology aside, the process behind the construction of audio deepfakes is a surprisingly simple one. Hackers have tweaked machine learning technology in such a way as to clone an individual’s voice, usually by utilizing spyware and devices that allow the cyber attacker to gather several hours of recordings of their victim speaking. The more data they are able to collect – and the better the quality of the recordings – the more accurate and potentially harmful the voice clone will be in practice. 

Once a voice model has been created, the malicious hacker’s AI gets to work ‘learning’ how to mimic the target. The AI will use what are known as generative adversarial networks (GAN), systems which continuously compete against one another through which one creates a fake and the other attempts to identify its flaws. 

With each new attempt, the AI is able to exponentially improve upon itself. This process continues until a reliable mimic is achieved and often succeeds after analyzing as few as twenty minutes of recordings.

[More to come ...]