Deepfakes mimic real people and pose serious risks. Learn how they’re made, their potential dangers, and the tools available for detecting them.
Amidst the breathless reports on the transformative power of artificial intelligence and the industry-altering new use cases arriving every day, there is a dark side. Like all paradigm shifts in technology, the mass adoption and accessibility of AI are making it possible for bad actors to execute increasingly sophisticated attacks, and chief among them is the use of deepfakes.
On the surface, most people think of deepfakes as the slightly off videos featuring celebrities or politicians doing something out of character. From Jim Carrey cast in the Shining to artificially generated news reports in Korea, deepfake technology is being used in new and creative ways to produce media portraying people and their voices in false ways. While they are most commonly related to actors, politicians, and public figures, deepfake technology is growing more sophisticated and capable of reproducing more people with less data.
How dangerous are deepfakes? Synthetic media has been used to create people out of whole cloth, create fake accounts, access real people’s financial accounts, and spread disinformation from otherwise trusted figures. And the technology is so new that many people don’t recognize these risks. Just last year, a survey by iProov indicated that 71% of respondents globally did not know what deepfakes were, and of those who did know what they were, only 57% said they thought they could spot one. Awareness is growing rapidly, but the skills and tools needed to spot deepfakes are not as accessible as many want. A Jumio survey found that 72% of consumers are worried about being scammed by AI technology, including depefakes, and an astounding 60% of consumers had encountered at least one AI-altered or fabricated video in the last year.
Whether deepfake images of the fires in Hawaii or a startlingly realistic portrayal of Morgan Freeman on social media, deepfakes are spreading rapidly, and consumers and businesses alike must have the resources to spot them and react accordingly.
Also Read: Deepfakes, Fraudsters, and Hackers Target Cybersecurity Jobs
How Deepfakes Are Made
Deepfakes are produced using deep learning ML algorithms. Using a neural network that emulates the complex connections of the human brain, deep learning ingests data in the form of images, videos, and audio, processes it, and produces fake output that emulates the properties of the input.
To ensure greater accuracy, there is a second algorithm that is trained to detect when an image or voice is faked. These two algorithms then work in tandem to produce, analyze, and dissect fake images to a higher degree of accuracy. This technology was once expensive and required large volumes of data to produce accurate fake images and audio, but it is now available commercially both online and as mobile apps. Nearly anyone can upload data into such an app and produce deepfakes, though the most sophisticated still tend to be produced by skilled operators.
The Risks Posed by Deepfakes
There are substantial risks associated with deepfakes. Perhaps most at the top of my mind is the impact of deepfake technology on political misinformation and the impact it could have on the 2024 elections. Rana Gujral recently joined Hanah Darley from Darktrace to discuss how deepfakes impact existing cognitive biases and voter perceptions and the significant impact that the boom in these attacks and fakes has on voter disinformation.
In 2023, fake newscasts circulated in social media with pro-China messaging created by deepfake technology. Later that year, a purported video of Volodymyr Zelensky showed him surrendering to Russia. Earlier this year, a Democratic political consultant was fined $6 million by the Federal Communications Commission and charged with several felonies in New Hampshire for creating an AI-generated voice call using a deep fake of Joe Biden’s voice, discouraging people from voting in the upcoming primary in that state. Misinformation, specifically related to political situations, is growing rapidly and adds to the melange of disinformation on social media and other online sources.
Another major area of concern is privacy and security. If deepfakes can replicate anyone with enough data to train an algorithm, how do we avoid widespread identity fraud and unlawful account access? A fraudster recently tricked a finance worker in Hong Kong in a video call with a deepfake of the company’s CFO and several other executives and stakeholders. Despite questioning a suspicious email, the people on the call all looked and acted familiar. The result was a loss of $25.6 million by the company.
Smaller instances of deepfakes are also being used, combined with classic social engineering hacks, to steal personally identifying information, hack into people’s accounts, and steal from them. And it’s not just video. A recent New Yorker article profiled the heartbreaking way in which voice cloning is being used to trick relatives into thinking their friends and family have been kidnapped, are in the hospital, or otherwise need immediate financial support.
From 2022 to 2023, the number of AI-powered deepfake frauds surged 10x globally, with the largest increase of 17x in North America. The rapid growth of these types of fraud show greater accessibility, ease of use, and affordability of deep fake images and voice cloning technology, which are increasingly effective with less data.
Also Read: Deepfakes Gone Wild: AI Threat or Creative Frontier?
Spotting a Deepfake
There are several cues that can help identify deepfaked images. These include facial and body movement that don’t look fully natural, small inaccuracies in lip syncing, and irregular shadowing on faces.
Audio deepfakes have artifacting introduced in the generated signal, and often don’t fully match the liveness of a human speaker. Both of these factors, however, can be difficult to identify without machine support. Liveness is subjective and relies on the context of the situation—if the cloned voice is stating an emergency, is it possible to take the time to assess breathing patterns or intonations to detect a deepfake?
For this reason, AI tools are needed to help detect faked voice and images. Artifacting, in particular, requires algorithms and models designed specifically to detect those signals. While voice-generating algorithms are growing more sophisticated detection technology relies on source analysis algorithms to analyze file metadata, analyze background details to find granular changes and assess signatures of existing content by training ML models using existing and previous deepfakes. Many of these tools have patterns of their own when creating deepfakes that can be, in turn, detected.
The Solution to Deepfake Deception
We are in the midst of a technology arms race, between fraudsters and those who would detect that fraud. In the short term, consumers are encouraged to be more vigilant and to question the source of media that appears in curated locations such as a social media feed, in an email, or over a phone call.
Behavioral Signals offers a personalized approach to deepfake detection that relies on an innovative three-stage process. Language agnostic behavioral profiling generates unique profiles for each speaker in a conversation, capturing speech patterns and characteristics based on factors like intonation, cadence, breathing, emotion, and more. By analyzing the nuanced factors of an individual’s speaking patterns, the technology can help ensure accurate detection of deepfake audio. Combined with speaker-specific detection, Behavioral Signals measures for anomalies to detect inconsistencies that indicate a potential deepfake, regardless of context or language.
This advanced approach to voice detection and analysis, allows organizations to identify anomalous communications, effectively monitor and analyze communications with foreign adversaries, authenticate voices before conversations take place, and generally enhance overall threat detection and protect critical assets.
Also Read: From Hacker to Defender: Sergey Belov’s Cybersecurity Playbook
The future of deepfake detection will be a combination of advanced technology and human vigilance. By better understanding how these fakes are made, we can respond to them in kind, reducing risk and improving protection for citizens, corporations, and governments.
