As you may have seen over the Christmas period, Channel 4 released their own version of the BBC’s Queen’s Christmas speech.

A tradition dating back to 1932 when King George V gave the first Christmas speech over radio, the Channel 4 speech last year was a little different in that the Queen did not appear anywhere in the video.

The message conveyed also came across as inconsistent from that of previous years. This is because it was a deep fake video in which the Queen’s likeness was replaced by falsified portrayal; just one example of open source software making a big impact on our mainstream media.

Open source software is a type of software in which the source code has been made freely available under a given open source license. But you may also be wondering what this has to do with deep fakes or the Queen? Well the process to create deepfakes in videos and/or photos using somebody else’s likeness is growing in common usage.

Creating deepfakes is easier than ever

The technology and techniques used by deepfakes is now readily available and easy enough for a person to create videos of anyone, providing they have sufficient source material.

Many of today’s popular apps, for instance, can work in real-time to swap your face with a selected static image. Using a pre-built machine model to identify and swap facial features like eyes, mouth and nose, it then uses those features as anchor points to map your face’s natural movements onto the static image.

Nevertheless, its effectiveness is limited to a tightly confined range of movements. Many of these apps struggle with large or rapid mouth movements due, in part, to the delay or frames per second (FPS) with which it’s able to process the image and apply adjustments.

Many deepfake applications also work from a single source image, and don’t require a slow training process for the program to “learn” the user’s facial features or structures. The quality of the results produced are quite varied between different images, and actors with wildly different looking faces.

While such deepfake apps present an interesting approach for showcasing the technology or even integrating it into a Zoom meeting, it’s unlikely to hold up to much scrutiny currently.

AI, machine learning and deepfake tech

So how about more complete solutions like DeepFaceLab? This follows a standard artificial intelligence (AI) and machine learning (ML) approach involving a training process to teach the algorithm to identify facial features from a series of still images.

First, these stills are extracted from video clips of the target and source videos. The images then need to be manually pruned to remove anything which may hinder the training process (i.e., other faces/elements of the images which aren’t faces). From there, the user can manually apply a mask to some of those images to define the face boundaries and features.

Next, it’s time to set the training task off and let the tool build its model. Once complete, the user can then use the model to apply the target face onto the source video.

While this tool is more complex to use, it also provides significantly better results. And although you don’t need to be an expert in video editing to use it, having some knowledge and access to such tools may further expand its promise.

Recommended

• Adopt cloud to fight climate change says UK Gov
• New and challenger banks top customer satisfaction survey
• Glasgow Uni spinout will create the first eco friendly insecticide

While open source can drastically increase the time to market and functionalities in software, it’s critically important to know what open source components have been included in each application.

As AI and ML applications continue to evolve and mature, it’s important to ensure open source code is used and maintained in a reasonable manner.

Let’s dig a bit deeper into the Zoom example. What could happen if the AI tools used by Zoom (and other video conferencing systems) had a critical bug? In other words, what if it’s vulnerable to a Remote Code Execution attack or is discreetly recording all voice or video discussions?

AI and ML aren’t unique types of open source so any vulnerabilities present in associated libraries can have serious wide-ranging impacts.

The difference, however, is that AI and ML are often viewed as a complex and difficult algorithm to understand, resulting in developers often having to rely on trust in the components they are using. But trust should not be given blindly.

Ensuring the components in use are free from security defects or that vulnerabilities are being addressed by developers helps to build and validate trust.

Deepfaking and trust

As AI and ML continue to shape more of our lives, and their impacts and consequences of failure grow, it’s going to become even more important to not only build AI that helps solve complex problems, but also justifies how it reached the conclusion it did.

As the 2021 OSSRA report shows, 84% of analysed codebases had at least one known vulnerability. This is further supported by more sophisticated attacks in which hackers have demonstrated the ability to intentionally add vulnerabilities to open source projects posing as helpful contributors.

Remaining aware and being an active member of the open source community is a valuable step to preventing these types of issues from occurring.

However, it’s unrealistic to expect every developer or company to be able to maintain this awareness on their own.

At a minimum, users of open source projects should consider performing regular reviews of the open source components they use and apply updates or patches when available.

It’s important to ensure that open source code is kept up to date with security patches so as not to be deceived by intruders –illustrating the intersection between open source and deepfake technologies.

Many projects build on common open source frameworks and technologies widely in use today, such as facial recognition and AI machine learning.

Open source is being applied in applications across all industries and is widely recognised as a substantial asset to developers and companies looking to re-define what is possible.

The future of open-source deepfakes

From keyless cars, recognising and unlocking themselves when the owner approaches, to re-creating beloved characters in your favourite TV shows – open source code provides the steppingstones to solve complex and seemingly impossible or impractical problems.

At the same time, open source comes with its own set of risks which need to be managed appropriately. While one of the many benefits of open source is its open nature, that also presents a variety of security vulnerabilities.

Sure, deepfake and the underlying AI and ML technologies are widespread buzzwords, but something often overlooked is that they’re based on open source software.

Rather than trying to rely on AI or ML to solve industry and market problems, first take a step back and consider whether the open source community has already solved it for you; just don’t forget to keep a vigilant eye out for vulnerabilities too.