From Entertainment to Exploitation: Deepfakes Threaten Truth In The Digital Age

Deepfakes are digitally produced images—similar to cinematic special effects—that enable fraudster individuals to generate realistic images and videos, which can be used to compromise biometric security systems. As the technology for creating deepfakes becomes more readily available, worries regarding fraud are on the rise. The potential threat posed by deepfakes is significant, and it is not just limited to entertainment on some social media websites; it can pose an imminent threat to the real world. To learn more about deepfakes technology, imminent threats, and the types of facial deepfakes, you can click the link: The Dark Side of AI: How Deepfakes Are Weaponizing Personal Identities?

We are here talking about deepfakes, which centre around the Generative Adversarial Networks (GANs) approach. We avoid cheapfakes, which are easy and quick to produce, especially tools for amateurs for entertainment purposes. Cheapfakes do not require GANs technology, sophisticated coding, or post-production skills. However, many experts think cheap fakes are dangerous because they are easy to produce with little effort.

Classes of Attacks on Biometric Systems

Biometrics utilise an individual’s distinctive physical or behavioural traits, like their facial features or voice, to identify and verify their identity remotely through their devices. Given the security and convenience issues associated with passwords, biometrics serve as an excellent alternative.

Biometric systems depend on the assurance that the user is physically present and active during the biometric data capture. In the absence of protective measures, fraudsters can successfully exploit this by employing non-live biometric representations; for instance, they can deceive facial recognition systems using images or videos that mimic the intended target, much like stealing a password.

There are principally two classes of attacks on biometric systems:

• Presentation attacks
• Injection attacks

Presentation Attacks

When a fraudster uses a victim’s physical characteristics or biometric data, such as paper photos or digital screens, to impersonate them, this is quite dangerous because malicious actors can use different presentation attack instruments, such as fake fingerprints or printed photos.

One way to mitigate presentation attacks is solid identity verification. A single identity verification can be operated on multiple layers, such as data sources, documents, and liveness. Advanced liveness verification can easily mitigate presentation attacks where advanced image processing is used to combat fraud. It includes image verification through pixel changes, deepfakes, and masks.

Injection Attacks

Injection attacks refer to hardware or software exploits that allow a perpetrator to “inject” digital images in a manner that circumvents the primary camera’s standard capturing process. This method can undermine liveness detection techniques that rely on an accurate capture process. Consequently, the fraudster can present their injected deepfake video as if it were genuine imagery recorded directly in front of the camera, even though it was not.

Let’s take a potentially threatening example. Let’s assume an influential political leader is giving a live speech, and fraudsters intercept the feed and replace it with a digital deepfake video. The audience sees their leader speak something unexpected and potentially dangerous.

With the improved efficacy of presentation attack detection, digital injection attacks are increasingly emerging as a common threat. These attacks involve introducing fake visuals at various attack points within a device, utilising either hardware or software techniques.

Mitigation of the Risk of Detection Attacks

Paper photos and screens are primarily used for presentation attacks, but deepfakes are totally different kinds of threats as they can be used for both presentation attacks and injection attacks. Therefore, only presentation attack detection (PAD) will not be sufficient for these kinds of attacks.

With the growing threat of deepfakes and the technological evolution of GANs, advanced countermeasures are implemented to mitigate these threats.

Multi-modal analysis: Multi-modal analysis involves AI detecting lip movement and mouth shapes to detect incongruities. The advanced multi-modal analysis system can detect even the minute difference between the original and Deepfakes words’ pronunciation.

Watermarking and digital signature techniques: This method uses an embedded digital signal or code within the video to identify the video’s origins.

Motion analysis: Motion analysis is a video analysis algorithm that detects discrepancies in movement in videos, such as a lack of motion blur.

Frame-by-frame analysis: Another form of video analysis involves analysing each video for inconsistencies, such as changes in lighting, shading, or texture.

Image artefact detectors: This is part of a machine learning algorithm specifically designed to detect Gans’s deepfakes. The algorithm recognises specific artefacts, such as non-symmetry of facial structure, teeth, background noise, etc.

Deepfakes Threat Detection: The S-L-T Approach

The S-L-T approach means a Socio-Legal-Technical framework that is implemented to mitigate the risks of deepfakes.

Societal (Regional)

• Raising awareness through media literacy
• Research funding through society funding to understand threat models and choice of infrastructure
• Organizational responsibility and accountability
• Prioritizing detection mechanism

Legal (National Level)

• Stringent government regulations and restrictions
• Content IDs and watermarking technology
• Regulations around dissemination of deepfakes

Technical (Global Level)

• Collaboration and partnerships
• Funding from the government, international forums, etc.
• Use of artificial intelligence technology to combat deepfakes

Conclusion

With the improvement of presentation attack detection methods, the threat of digital injection attacks is on the rise. Attackers can inject fake imagery at various locations within a device, utilising either hardware or software vulnerabilities.