Image commercially licensed from Unsplash
Deep Fakes, artificially generated videos or images that convincingly replace a person’s likeness and voice, have emerged as a disruptive force in the digital landscape. Initially touted for benign uses like video editing and animation, Deep Fakes are increasingly being weaponized for disinformation, identity theft, and cyberbullying. With technology evolving rapidly, the line between reality and synthetic media is becoming blurrier, making it imperative to develop robust detection and mitigation techniques.
Articles around the internet show some viral videos circulating on the web and social media. However, the disturbing use of deep fakes to scam people and identity theft raises a serious concern. Social media platforms such as TikTok and Instagram have seen a recent spike in video content imitating celebrities, prominent investors, etc., to attract funds for nonexistent businesses.
The news organization NBC News mentioned that they came across several dozens of videos posted to social media sites featuring computer-manipulated images and audio clips of prominent people like Elon Musk and Donald Trump, all of which appeared to have been created to scam viewers out of money.
It was shocking to see that a large majority of the videos were centered on Elon Musk, along with edited videos of news and television personalities — including CBS News anchor Gayle King, former Fox News host Tucker Carlson and HBO host Bill Maher — falsely claiming that Elon Musk had invented a technologically advanced investment platform. Elon Musk’s history of promoting cryptocurrency in the past adds to the pool of confusion.
Given that a plethora of almost free tools offer deep fake editing, it would be impractical to think of ways to stop deep flakes, but it would be wise to come up with ways to detect deep fakes and flag them.
Let’s discuss some traditional methods that can help detect deep fakes; while some of these methods may seem elementary, many of the videos on social media by amateur editors can be caught with such simple methods.
Traditional Methods for Deep Fake Detection
One of the earliest methods to detect deep fakes focused on visual inconsistencies; factors like inconsistent lighting, unnatural eye movements, or even irregular blinking rates could serve as red flags. However, as technology improves, these visual clues are becoming less reliable.
Sound engineering needs to catch up with visual manipulation in Deep Fake technology, making audio analysis a viable detection method. Discrepancies in voice texture, tone, and ambient noise can indicate manipulated media.
Every digital file comes with metadata that can provide clues about its origin. By analyzing this information, one can sometimes trace back to the software used for creating the Deep Fake, thereby flagging it for further review.
A few other advanced methods that can help are as follows:
Machine learning algorithms, particularly neural networks, have shown promise in Deep Fake detection. These algorithms are trained to recognize patterns and inconsistencies generally overlooked by the human eye.
Generative Adversarial Networks (GANs)
GANs, which ironically are often used to create Deep Fakes, can also be utilized for detection. A GAN-based detector is trained to distinguish between natural and synthetic media, continuously improving its accuracy through adversarial training.
Incorporating Facial Recognition Techniques
Face recognition plays a pivotal role in creating and detecting Deep Fakes. The paper by Bahmani et al., 2021, presents a novel method to quantify bias using skin reflectance as a measure that can improve facial recognition efficiency among various demographics. If a Deep Fake model fails to accurately mimic the skin tone reflectance patterns of the individual it aims to impersonate, advanced facial recognition algorithms could flag it as synthetic media.
Guarnera et al., 2020, focused on the analysis of Deepfakes of people’s faces to create a new detection method that may be able to detect a forensics trace hidden in pictures: a unique type of fingerprint retained from the image generation process. Their proposed technique, using an Expectation Maximization (EM) algorithm, extracts a set of local features specifically addressed to model the underlying convolutional generative process. These technologies could be leveraged for Deep Fake detection by improving the accuracy of biometric validation.
Dr. Sahu, a biometrics expert and co-author of the SREDS paper, says, “The use of skin tone in deep fake detection is a complex issue with no easy answers. On the one hand, skin tone can be a useful feature for identifying deep fakes, as it is often difficult to replicate the subtle variations in skin tone that occur naturally. On the other hand, the use of skin tones for deep fake detection can be discriminatory, as it could disproportionately impact people with darker skin tones.
There are a number of ways to address the discriminatory potential of skin tone-based deep fake detection. One approach is to use a combination of features, such as skin tones, facial structure, and eye movement, to make identifications. This can help to reduce the reliance on any single feature and make the identification process more robust. An alternative approach is to use machine learning algorithms that are trained on a diverse dataset of images, including images of people with all skin tones. This can help to ensure that the algorithm is not biased against any particular group of people.
Ultimately, the use of skin tone in deep fake detection is a trade-off between accuracy and fairness. There is no perfect solution, but by carefully considering the risks and benefits, it is possible to develop methods that are both effective and equitable.”
Blockchain for Content Verification
Blockchain technology could offer a solution for content verification. Maintaining a decentralized, immutable ledger of original content makes it easier to verify the authenticity of digital media.
Watermarking digital media with a unique, non-removable identifier could also be a mitigation strategy. This would make it easier to trace the origin of the content and confirm its authenticity.
Deep Fakes pose a complex challenge requiring a multifaceted detection and mitigation approach. Leveraging advanced methods like neural networks and incorporating novel facial recognition techniques, such as those discussed by Guarnera et al., 2020 and Bahmani et al., 2021, can significantly bolster our defense mechanisms against this emerging threat. As the technology behind Deep Fakes continues to evolve, so must our methods for detecting and mitigating them.