Deepfake Deception, Detection and Dismantlement
Deepfakes. A trend that took the internet by storm and enthralled hundreds of creators and viewers by watching and making their favourite actors and public figures say and do whatever they want. They have been around for a while now and the trend is still as popular as ever with DeepTomCruise being a hit on TikTok, an account that has deepfake imagery and videos of Tom Cruise doing pretty normal things, but unnervingly realistic and not actually him. As technology advances and people are learning so fast how to use these tools, it is becoming more difficult to tell real from fake. Therefore, Facebook has teamed up with Michigan State University and are taking on the challenge of developing a method of detecting a deepfake and then using reverse engineering to discover the machine learning model that created it.
The issue surrounding deepfakes isn’t really a cause for concern on Facebook currently, but there are plenty of negative implications these types of images and manufactured videos could have. Specifically when it comes to elections, voters may have their mind changed if they see one of these videos with their chosen candidate doing or saying something that wasn’t really them, which could damage the future of polls and how they will be able to manage campaigns if these false materials slip through the radar. Also, misinformation and false news stories could cause disruption in general, with a wide variety of this is already being spread around Facebook and other social media sites, faking news broadcasts could cause a stir in communities especially due to the pandemic.
There have been previous studies in this field and they have been able to determine which AI model generated the deepfake, but new research ,led by Michigan State University’s Vishal Asnani, is progressing this research by identifying architectural traits of models that are unknown. The traits are known as hyperparameters and need to be fine tuned in each machine learning model similar to parts in an engine, these leave a sort of fingerprint unique to themselves on the finished media which will be able to then identify the source. Being able to figure out which pictures come from the same device, it will mean that they will be able to seize the laptop or computer that produced the content. This research is the first-time people have been able to identify certain properties of the model used to create the content, without them having any prior knowledge of any specific model.
“If this is a new AI model nobody’s seen before, then there’s very little that we could have said about it in the past. Now, we’re able to say, ‘Look, the picture that was uploaded here, the picture that was uploaded there, all of them came from the same model.’ And if we were able to seize the laptop or computer [used to generate the content], we will be able to say, ‘This is the culprit.’” Says Facebook research lead Tal Hassner.
The algorithm can work out unique pointers about the generative model, but it can also help to identify which images are deepfakes, which is a leap, although detecting all deepfakes is still an unsolved problem, as the state of the art software isn’t always reliable. There are many attributes than the AI needs to work out and decipher how and why the image is manipulated. Facebook held a deepfake detection competition last year for developers, and algorithm that won could only detect the deepfake videos 65.18% of the time. Work for identifying and shutting down the creators of deepfake images is an ongoing study, but at some point, the code will be cracked.
As with all technology, knowledge of this subject is becoming more advanced every day, with new techniques on how to actually hide their computer fingerprints from the pictures and videos, but there are many a dedicated teams that are working hard to solve the issues and work around the chasing back and forth, although it may be a long time yet.
“This is a cat and mouse game, and it continues to be a cat and mouse game.” -Tal Hassner.
Keep up-to-date with the latest tech industry insights, trends as well as information technologies, app development, and small business content with the Proteams Blog
Follow us on LinkedIn for updates on the latest tech news here