Using Deepfake Detection AI to Combat AI Fraud in Applications

2023-05-18

An Overview of Deepfakes and How They Are Used for Financial Fraud Deepfakes have been a hot topic in the field of data science for many years. As early as 2020, MIT Technology Review stated that deepfakes had reached the "tipping point of mainstream use." Undoubtedly, the data confirms this. According to The Wall Street Journal, there were less than 10,000 deepfakes discovered online in 2018. Now, these numbers have reached millions, and there are many real-life examples that demonstrate how deepfakes are used to confuse and mislead information, as well as perpetuate financial fraud. Deepfake technology provides cybercriminals with many advanced options. They go far beyond the ability to insert celebrity images into "not-to-be-missed" Bitcoin promotions, which are, of course, scams. Deepfake videos, in particular, are vulnerable to fraudsters' attention. Deepfake videos provide them with a method to bypass ID and KYC checks, which has proven to be highly effective. In May 2022, The Verge reported that "liveness tests" used by banks and other institutions to help verify user identities are easily deceived by deepfakes. Research has found that 90% of identity verification systems have vulnerabilities. So, what is the answer? Are we entering an era where cybercriminals can easily exploit deepfake technology to bypass the security measures of financial institutions? Do these companies have to abandon their automated systems and resort to manual, human checks? The simple answer is "probably not." Just as criminals can leverage the rapid development of artificial intelligence, so can their target companies. Now, let's see how vulnerable companies are using artificial intelligence to fight against artificial intelligence. How do Deepfakes work? Deepfakes are created using a series of artificial intelligence techniques, such as: - Generative Adversarial Networks (GANs) - Encoder/Decoder - First-order motion models At first glance, these technologies may sound exclusive to the machine learning community, with a high barrier to entry and requiring specialized technical knowledge. However, just like other elements of AI, they have become increasingly accessible over time. Low-cost, off-the-shelf tools now allow non-technical users to create deepfakes, just as anyone can register with OpenAI and test the capabilities of ChatGPT. In 2020, the World Economic Forum reported that creating "state-of-the-art" deepfakes cost less than $30,000. But in 2023, Ethan Mollick, a professor at the Wharton School, revealed through a viral tweet that he made a deepfake video of himself giving a speech in just 6 minutes, costing him $10.99. He used a service called ElevenLabs to mimic his voice almost perfectly for $5 and another service called D-ID, which generated the video based on a script and a single photo for a monthly fee of $5.99. He even used ChatGPT to create the script. When deepfakes first emerged, the main focus was on fake political videos (and fake pornography). Since then, the world has witnessed: - BuzzFeedVideos creating a deepfake public service announcement featuring actor Jordon Peele impersonating Barack Obama. - A deepfake YouTube video claiming to show Donald Trump telling a story about reindeer. - Saturday Night Live airing a deepfake video of Hilary Clinton, who was actually portrayed by an actor. While these examples showcase the "fun" side of deepfakes and perhaps provide some realistic understanding of the technology's capabilities, fraudsters wasted no time in using it for malicious purposes. There are numerous real-life examples of fraud using deepfake technology, resulting in losses ranging from hundreds of thousands to millions of dollars. In 2021, an AI voice cloning scam was used to orchestrate a $35 million fraudulent bank transfer. This was a significant economic gain without even needing to use video. The quality of AI-generated output, especially videos, can vary greatly. Some videos are obviously fake to humans. However, as mentioned earlier, past evidence has shown that the automated systems used by banks and fintech companies are easily fooled. As AI capabilities continue to improve, this balance may further shift. A recent development is the combination of "counter-forensics" and adding "targeted imperceptible" noise to deepfakes in an attempt to deceive detection mechanisms. So, what can we do? Using AI to Combat AI: Detecting Deepfake Fraud While fraudsters attempt to exploit the latest AI technologies for economic gain, technology companies and other enterprises are also striving to find ways to catch criminals using technology. Here are some examples of companies using AI to combat AI: - In late 2022, Intel launched an AI-based tool called "FakeCatcher." According to reports, Intel reported a reliability rate of 96% using a technique called photoplethysmography (PPG). This technology leverages something that is absent in artificially generated videos: blood flow. The deep learning algorithm, trained on legitimate videos, measures the light absorbed or reflected by blood vessels, as the color of the vessels changes when blood flows through different parts of the body. FakeCatcher is part of Intel's Responsible AI program and is described as the "world's first real-time deepfake detector, returning results in milliseconds." It is an innovative technology that looks for signs of whether the person depicted in the video is a real human. It looks for what is "right" rather than highlighting what is "wrong" through data analysis. This is how it indicates the likelihood of a forgery. - Meanwhile, computer scientists at the University at Buffalo have been researching their own deepfake detection technology. It uses something known to avid PC gamers, which requires significant processing power to emulate: light. The University at Buffalo claims that this AI tool can observe the reflection of light in the eyes of the subject in the photo, achieving a detection rate of 94% for fake photos. The surface of the cornea acts as a mirror and produces a "reflection pattern." This suggests that imitating genuine highlights with AI systems will be no small feat. PC gamers who invest in the latest ray-tracing graphics cards for the sake of realistic lighting effects will instinctively recognize the challenge here. The future of AI in fraud detection Over time, the nature of machine learning determines that it will become better at detecting anomalies and combating fraud. AI-driven systems can learn from new patterns and potentially filter out fraudulent transactions at an early stage of the process. When it comes to deepfakes, the examples above give people a particular hope. Scientists have found a way to detect the majority of deepfakes by utilizing light reflection. Developments like this represent a significant step forward in preventing fraud and pose a major obstacle for cybercriminals. In theory, deploying such detection technology is much easier than fraudsters finding ways to bypass it, such as replicating the behavior of light at speed and scale. The "cat and mouse" game may seem like it will continue indefinitely, but at least in theory, large tech and financial companies with resources and substantial funding can stay one step ahead. However, the existence of such mechanisms is one thing, but seeing them frequently integrated into solutions used by enterprises is another. Earlier, we mentioned a statistic that indicated 90% of solutions can be "easily fooled." There may still be some financial institutions using such systems. A wise fraud detection strategy requires companies to look beyond detecting deepfakes themselves. There is much that can be done before fraudsters get involved in video-based ID verification or KYC processes. Preventive measures that identify their presence earlier in the process may also involve elements of artificial intelligence and machine learning. For example, machine learning can be used for real-time fraud monitoring as well as creating rule sets. These can look for patterns that humans are prone to miss by examining the history of fraud incidents. Transactions deemed high-risk can be directly rejected or flagged for manual review before potentially undergoing ID checks, thus minimizing the opportunity for fraudsters to exploit deepfake technology. The sooner the system detects cybercriminals, the less likely they are to continue their criminal activities, and the less cost companies incur in further investigations. Even without incorporating AI technology to detect deepfakes, the cost of video-based ID checks is high. If fraudsters can be identified before they reach that stage, utilizing techniques such as digital footprints, more resources will be available to optimize checks on more marginal cases. The future of AI in fraud detection holds promise. As machine learning continues to evolve, it will become better at detecting anomalies and combating fraud. The ability to detect deepfakes by leveraging light reflection represents a significant step forward in preventing fraud and poses a major obstacle for cybercriminals. While the "cat and mouse" game between fraudsters and those dedicated to thwarting them may continue indefinitely, large tech and financial companies with resources and substantial funding can maintain a slight advantage. Deploying detection technology is theoretically easier than fraudsters finding ways to bypass it. However, it is crucial for companies to focus not only on detecting deepfakes but also on implementing preventive measures earlier in the process to minimize the opportunities for fraudsters to exploit deepfake technology. In conclusion, AI-powered fraud detection strategies, combined with ongoing advancements in technology, offer hope in the fight against deepfake fraud. By staying vigilant and continuously improving detection methods, companies can mitigate the risks posed by deepfakes and protect themselves and their customers from financial fraud.

RECENT AI NEWS

RECENT AI TOOLS