In the Era of Deepfake Proliferation, Can We Still Believe What We See?

Deepfake technology is garnering widespread attention. The term originates from "deep learning" and "fake," primarily referring to using artificial intelligence deep learning to synthesize images, videos, or even audio that may not necessarily exist in reality. The most common application is swapping one person's face for another's in a video. With Deepfake-related technology, many things that previously had to be abandoned due to regret can now be compensated for. The most famous applications are found in the Hollywood films Fast & Furious 7 and the Star Wars series. Paul William Walker IV, a lead actor in Fast & Furious 7, passed away unexpectedly during filming. However, the production team utilized Deepfake technology, using Paul's brother as a stand-in and performing digital facial synthesis, allowing the film's production to continue and be released successfully. Such technology provides solutions to previously unsolvable problems, demonstrating the application value and possibilities of Deepfake technology.

Traditional software algorithms primarily rely on continuous modification and adjustment by engineers; however, technologies like Deepfake use internal algorithms that are trained and evolve continuously.

When discussing how to improve the deceptive effects of Deepfake technology, a variety of techniques and methods are widely used, one of the most common and representative being the "Generative Adversarial Network (GAN)." This technical framework includes two key neural network components: the "Generator" and the "Discriminator." These two act like a pair of mutually influential characters, much like a tiny angel and a tiny demon sitting on a master's soul, exchanging opinions, evolving together, and constantly improving their respective skills and capabilities.

Initially, when the training process begins, these two sets of neural networks are in a state of ignorance, much like newborn babies. However, as time passes and experience accumulates, they gradually become smarter. This training process can be compared to a sophisticated debate where the "Generator" and the "Discriminator" compete against each other, constantly trying to propose more persuasive arguments and improving each other's skills.

The "Generator's" job is to continuously produce forged images, while the "Discriminator's" task is to evaluate the authenticity of these images and assign them scores. Over time, the "Generator" becomes increasingly skilled at creating realistic forged images because it must constantly chase the threshold set by the "Discriminator," making it difficult to distinguish truth from falsehood.

This adversarial process is like an endless training session; it is crucial for improving the technology of forged image generation and simultaneously drives the development of anti-counterfeiting technology, making it more powerful and intelligent. This balance between competition and collaboration allows Deepfake technology to progress continuously, constantly challenging the boundary between identifying the real and the fake.

An increasing number of mobile applications and open-source software now have these features built-in, allowing the general public to easily access and use these technologies. Through simple operations, multimedia content editing that traditionally required complex steps can now be performed. As a result, the barrier to entry for deepfake technology has been significantly lowered, and it can be used by malicious actors for the mass production of forged celebrity pornography, similar to the influencer face-swapping incident that drew widespread social attention late last year.

Through Deepfake technology, we can easily replace a face in a video with anyone we want and simulate their specific expressions and movements. This technology can also be used to create fake news, forge accounts, and conduct fraudulent activities. For example, Facebook and Instagram have repeatedly discovered that many accounts, fan pages, and groups using deepfake-generated photos were being used for scams or to spread sensitive rhetoric. A real case involves the CEO of a British energy company who received a call from his superior at the German parent company requesting a transfer of 7.8 million TWD to a designated supplier account. The CEO, believing he was very familiar with his superior's voice, executed the transfer without hesitation (News Source, January 2019). In reality, the superior's voice was forged through Voice Conversion technology. As new deepfake models continue to emerge, fraudulent acts are occurring incessantly. Although some generated images and audio are of poor quality, users still find it difficult to determine if content has been modified due to the limited size and resolution of mobile phone screens. Traditional visual evidence can no longer guarantee authenticity 100%, necessitating a deeper understanding of deepfake-related issues across all sectors of society.