Photo credit: Tima Miroshnichenko
Deepfakes have been a hot topic in the data science community for the last few years. Back in 2020, MIT Technology Review claimed that deep fakes had reached a “tipping point in mainstream usage.”
The data certainly back it up.of wall street journal reported that fewer than 10,000 deepfakes were discovered online in 2018. That number now stands in the millions, and there are many real-world examples of deepfakes being used to confuse, misinform, and perpetuate financial fraud.
Deepfake technology offers cybercriminals many advanced options.
They go far beyond their ability to insert celebrity images into promotional materials for their “can’t miss” Bitcoin offer, which, of course, turns out to be a scam. Deepfake videos in particular have attracted the attention of scammers. They provide a way to get past automated ID and KYC checks and have proven to be terrifyingly effective.
In May 2022, The Verge It reported that “survival tests” used by banks and other institutions to verify a user’s identity can be easily fooled by deepfakes. A related study found that 90% of tested identity verification systems are vulnerable.
So what would be the answer? Is it time for cybercriminals to use deepfake technology to easily bypass financial institutions’ security measures? Should these companies abandon automated systems and return to manual human checks?
The short answer is “probably not”. Just as criminals can take advantage of her rapid advances in AI, so can targeted companies. Now let’s see how vulnerable companies can use AI to combat AI.
Deepfakes are created using various artificial intelligence techniques such as:
- Generative Adversarial Network (GAN)
- Encoder/decoder pair
- Primary motion model
At first glance, these techniques may sound like a proprietary area of the machine learning community, but they have high barriers to entry and require specialized technical knowledge. However, like other elements of AI, they have become much more accessible over time.
Just as anyone can sign up for OpenAI to test ChatGPT’s capabilities, low-cost, off-the-shelf tools can now be used by non-technical users to create deep fakes.
In 2020, the World Economic Forum reported that “state-of-the-art” deepfakes cost less than $30,000 to produce. But in 2023, Wharton School professor Ethan Morrick revealed through his viral Twitter post that he had created a deepfake video of himself giving a lecture in less than six minutes.
Morrick’s total spending was $10.99. He used his service, 11Labs, to mimic his voice almost perfectly for $5. Another service called D-ID, for $5.99 a month, generated a video based on a script and just one photo of him. He used his ChatGPT to create the script itself.
When deepfakes first started to emerge, the main focus was on fake political videos (and fake pornography). Since then, the world has seen:
- BuzzFeedVideos will create a deepfake public service announcement “featuring” Barack Obama impersonating actor Jordon Peele.
- A deepfake YouTube video purporting to show Donald Trump talking about reindeer.
- Deepfake video of Hillary Clinton aired on Saturday Night Live, where she was actually impersonated as a performer.
While these examples show the “fun” side of deepfakes and perhaps some real-life shocks about how this technology works, scammers are wasting their time using deepfakes for their nefarious purposes. Is not …
There are many examples of frauds perpetuated using deepfake technology.
Losses from deepfake scams range from hundreds of thousands to millions. In 2021, he was arranged for $35 million in fraudulent bank transfers using an AI voice clone scam. This brought enormous economic benefits, Required Use of video.
AI output, especially video quality, can vary greatly. Some videos are clearly fake to humans. However, as mentioned above, the automated systems used by banks and fintechs have proven to be easily fooled in the past.
The balance could shift further as AI capabilities continue to improve. Recent developments have incorporated ‘counterforensics’, which adds ‘targeted and invisible ‘noise’ to deepfakes in order to deceive detection mechanisms.
What can you do with it?
Just as fraudsters seek to use the latest AI technologies for financial gain, tech companies and other companies are also working hard to find ways to use technology to catch criminals. is.
Here are some examples of companies using AI to fight AI.
In late 2022, Intel announced an AI-based tool called “FakeCatcher.” Intel reports a reliability rate of 96% and uses a technology known as photoplethysmography (PPG).
This technology takes advantage of something that isn’t present in artificially generated videos: blood flow. Its deep learning algorithm, trained on regular videos, measures light absorbed or reflected by blood vessels, which change color as blood moves through the body.
Part of Intel’s Responsible AI initiative, FakeCatcher is described as “the world’s first real-time deep fake detector that returns results in milliseconds.” This is a revolutionary technology that looks for signs that the person in the video is really human. Instead of analyzing the data and emphasizing what is “wrong”, look for what is “right”. This way it shows the possibility of fake.
Meanwhile, computer scientists at the University of Buffalo (UB) are working on their own deepfake detection technology. It uses something that avid PC gamers know requires a huge amount of processing power to emulate: light.
Claimed by UB to be 94% effective on fake photos, this AI tool looks at how light reflects off a subject’s eyes. The corneal surface acts as a mirror and produces a ‘reflection pattern’.
Entitled “GAN-Generated Face Exposure with Unmatched Corneal Specular Highlights,” the scientist’s study states that “GAN-synthesized faces are exposed with unmatched corneal specular highlights between two eyes.” It shows that you can.
It suggests that it is “non-trivial” for an AI system to emulate real highlights. PC gamers who often invest in the latest ray-traced graphics cards to experience realistic lighting effects will intuitively recognize the challenges here.
Perhaps the biggest challenge in fraud detection is the never-ending “cat-and-cat game” between fraudsters and those trying to stop them. Following announcements like the one above, it is very likely that people are already working on building technologies that can circumvent and defeat such detection mechanisms.
Also, the existence of such a mechanism is one thing, and the fact that it is routinely integrated into the solutions companies use is one thing. Earlier I mentioned a statistic that suggests that 90% of solutions can be “easily fooled.” At least some financial institutions may still use such systems.
A smart fraud surveillance strategy requires companies to look beyond detecting deepfakes themselves.there’s a lot you can do Before Scammers penetrate deep into systems and participate in video-based identity verification or KYC processes. Precautionary measures to locate locations early in the process may also include elements of AI and machine learning.
For example, machine learning can be used for both real-time fraud monitoring and ruleset creation. They can examine past misconduct and detect patterns that humans might easily miss. Transactions deemed high risk may be rejected outright or passed to manual review. before reaching It is a stage where identity checks can occur, thus presenting an opportunity for fraudsters to take advantage of deepfake technology.
The sooner the system detects cybercriminals, the better. Crime is less likely to be perpetuated, and businesses spend less on further inspections. Video-based identity checks are costly, even if they don’t incorporate AI technology to detect deepfakes.
Using techniques such as digital footprinting to identify fraudsters before they get there will free up more resources to optimize the checking of more borderline cases.
The very nature of machine learning should make it better at detecting anomalies and fighting fraud over time. AI-powered systems could learn from new patterns and eliminate fraudulent transactions early in the process.
Especially when it comes to deepfakes, the example above gives us special reason to be hopeful. Scientists have found a way to detect the majority of deepfakes using reflected light. This development is both a major advance in fraud prevention and a major obstacle for cybercriminals.
In theory, it would be much easier for a fraudster to deploy such detection technology than to find a way around it, for example replicating the behavior of light at high speed and at scale. The cat-and-mouse game seems to go on forever, but big tech companies and big financial institutions have the resources and deep pockets to stay ahead, at least in theory.
Jimmy Fong He is the CCO of SEON and brings deep experience in fighting fraud to assist fraud teams everywhere.
