“Deepfakes” are back in the news after a software called “deep-live cam” was released on Github. Users are circulating photos and videos that impersonate political pundit J.D. Vance using this new technology. What makes this software package different from the deep-fake videos of the past is the ability for face swap to occur in real time. Previously, users could only alter faces or voices in post-production. Now, someone using the software can face-swap while they are filming. To make matters worse, Github is an open-source site which means any interested party can download and use this advanced impersonation software for free.
Stay tuned as we delve into how this technology works and the implications.
What is a "Deepfake"?
A deepfake is a photo or video that uses artificial intelligence software to show any desired individual in any setting according to the prompts entered by the user. The term originated from Reddit in 2017 with “deep” referencing deep learning and “fake” being not real (we will elaborate more on deep learning later in the article.) Deepfakes are arguably the most visible examples of artificial intelligence’s entrance into the public landscape.
One of the unfortunate realities of this software is that users mainly employ it for malicious purposes, including misinformation campaigns and pornography. At its most benign, users create deepfakes for comedy and satire. At its worst, deepfakes ruin people’s reputations by imposing their likenesses onto graphic content without their consent.
With Github’s latest release, the reaches of deepfake software is even greater, and thus has greater consequences.
What Makes Deepfake Live Different
Deepfake Live stands apart from previous versions of generative AI software because of its ability to replicate faces and voices in real-time. Like a filter on Instagram or Snapchat, the imposed face can move with you, compensate for lighting changes, and more using only one photo as a reference. The accuracy is startling. Paired with existing voice cloning technology, the implications are grim. Users have been able to scam family members and multi-national financial workers alike using deepfake video and phone calls. AI-powered fraud is now one of the top-5 forms of identity fraud with a 1740% surge in deepfakes in North America from 2022 to 2023.
How Do Scammers Use Deepfakes?
1. IMPERSONATING PEOPLE YOU KNOW
Extortion using Deepfake technology can take many forms. In one of the most extreme cases, a group of scammers tricked a financial worker into paying $25 million. The scammers were using Deepfake video and audio software to impersonate the COO of the employee’s company and other coworkers on a video call. Other instances include grandparents receiving calls from “their grandkids,” “friends” needing bail money, and similar emergencies that leave victims in a high-urgency situation with someone they would not think to question or verify.
Scammers often call the individuals they plan to impersonate, in order to compile data on their financial and social networks as well as their vocal intonations. Then, the fraudsters find a victim and use Deepfake audio and/or video software to perform the scam call. This can happen over video, as seen in the example of the COO impersonation described earlier, or even just over the telephone.
2. EXPLOITING TRUST IN PUBLIC FIGURES
Another use is the fabrication of “celebrity endorsements” for crypto scams. Generative AI processes videos or TV clips to alter the faces, audio, and movements of the original speaker. There is often an uncanny nature to these videos and a viewer might be able to detect them as AI, but someone not looking closely could easily miss the signs.
Fake interviews by Tucker Carlson and speeches by Elon Musk have deceived prospective investors out of hundreds of dollars. The con artists behind the Quantum AI scam, a “get rich quick” investment scheme, have used these tactics. Research by the Threat Research Center at Palo Alto Networks reveals that these scam sites appear to stem from a single source. The domain’s victims are funneled to repeat despite having different sources and share a DNS provider. Additionally, the scam websites use the same videos to promote the crypto scam across different sites.
The Technology Behind Deepfake-Live
Let’s get into the “deep” in Deepfake. Deep Learning is a more extensive version of Machine Learning—one of the foundational components of artificial intelligence. Machine learning is what grooms AI into imitating human neural networks. The only difference between Machine Learning and Deep Learning is the quantity of computation layers involved in processing data. There are typically two computation layers in Machine Learning, while Deep Learning involves 3 or more, though it can often stretch into the hundreds and thousands.
What is a computation layer? Think of a layer as a filter; it is a means of processing information. Just like humans process and react to stimuli, these computational layers imitate those same functions. In reference to Deepfakes and Deepfake-Live, the main layers at play are the Convolutional layers. These are the “eyes” and “ears” of generative AI. They are responsible for processing images and audio. After inputting an image into AI software, convolutional layers (typically referred to as CNNs) analyze the edges, textures, and patterns of the image and convert that information into a grid-like topology. Then, each data point on the grid is translated into binary code. Once this code has been processed by the AI software, the topology can be projected onto another object in real-time.
audio cloning
When it comes to voice cloning, the CNNs reproduce audio patterns in a process similar to a vocoder. The only difference is this “vocoder” replicates sounds based on AI’s interpretation of audio patterns rather than a user’s manual configuration or audio recording data. Audio is processed traditionally by snapshotting the numeric values along a sonic wavelength. The number of snapshots taken depends on the frame rate of an audio recording. This is amplitude x time waveforms.
CNNs, on the other hand, process audio using data in the format of spectrograms. Spectrograms plot time on the x-axis and frequency on the y. Then, these individual data points are colored to correspond with the strength of signal being transmitted (the amplitude). This allows spectrograms to represent not only the strength of an audio file over time, but the frequency (pitch range) of said audio. This is a process of deep learning and is what allows for “deep fakes” to be more realistic. The more data provided, the more accurate the reproduction.
Even just a brief summary of the technology behind Deepfake-Live shows its complexity. With new information and software being released on open-source sites like Github, the capabilities and reach of generative AI advances with every passing day.
What to Do in the Age of Deepfake-Live
PROCEED WITH CAUTION
As with every scam these days, caution is key. Modern tech users are familiar with well-established means of phishing like email, phone, and texting scams. Now, the landscape has evolved such that users should be wary of even phone or video calls from a loved one. Whether it’s establishing a safe word with your family members or a set of baselines for sensitive communications within your company, you can never be too safe when it comes to protecting your data and finances. Needless to say, never pay someone or invest in something unless you are sure it is legitimate. The method of payment requested by the scammer is often a dead-giveaway: Cryptocurrency, wire transfers, gift cards, and bank account information are the preferred methods for fraudulent transactions.
PROTECT YOUR DATA
You can also take some precautions to protect your likeness. Encrypting your data whenever possible is one safeguard against Deep Learning. AI cannot process encrypted image data into a Convolutional layer. Encryption acts as a gate protecting the metadata. On iOS there is an “advanced data protection” setting localized on iPhones. Android does not appear to have a similar localized method of encrypting or protecting data. Windows devices and cloud storage are end-to-end encrypted, iCloud is not. For data on Android or Mac that you might want encrypted or protected, a third-party storage or file sharing app might be necessary.
Potentially exploitable metadata from social media sites is unfortunately much easier to access. Sites like Instagram and Facebook do not encrypt the photos and videos we post to them. Meta offers end-to-end encryption in Messenger, while WhatsApp encrypts photos, audio, and text sent within its network. Thankfully, at least the private messages we send are less likely to have their metadata exploited. However, this does not account for all the countless other ways we share content online.
Always Double-Check
In this day and age, it can be extremely difficult to discern what is real and what is fake. If something seems too good to be true, is out of character, or just feels fishy, it is always good to investigate further. If there is money on the line, always proceed with extreme caution. Chances are the boss of your company is not calling you in a dire situation needing a few thousand dollars. Always double check because otherwise, the outcome could be disastrous.
