If one has read the newspaper ‘The Australian’, they must have read how a machine generated image was used as a synthetic warrior to wage cyber warfare. In other words, they used a deep fake.
Lets understand a little bit on
- What is a Deepfake and how did it start?
- How does one make Deepfake?
- How do you suss out a Deepfake?
- How rogue can it get?
What is a Deepfake and how did it start?
First the genesis of the word ‘deepfake’ – it is a combination of ‘deep learning’ and the word ‘fake’.
Deep learning, as we all know, rolls up to machine learning, which in turn rolls up to Artificial Intelligence. Deepfake is through how one can create unreal images, videos or audio recordings. Increased computing power and the accompanying ability to process and analyze massive amounts of data has enabled us to synthesize images, videos and audios of events which never occurred.
Maybe it all started with this Video Rewrite paper published by Purdue University back in 1997. To quote verbatim from the paper ‘Video Rewrite uses existing footage to create automatically new video of a person mouthing words that she did not speak in the original footage. This technique is useful in movie dubbing, for example, where the movie sequence can be modified to sync the actors’ lip motions to the new soundtrack’.
About 20 years later, in 2017, the term ‘deepfake’ may have been coined and occurred first on the reddit platform who used the technology for really unsavory purposes. Because of the nature of the use case for which it was used, it may have gone viral too fast, catching the attention of bad actors, morphing into a tool to used for nefarious activities
How does one make Deepfake?
Without going in detail about how to make a Deepfake – we can try to understand in general how a face is switched. The first step would be for an encoder to work on 2 different faces – reducing them to a common set of compressed features. The second step would be for a decoder to work on learning how to reconstruct the original faces from the compressed feature. The third step would be to swap or in other words feed the ‘other face’ compressed set of features to the decoder, which will re-construct the wrong face on the right body – or the right face on the wrong body – however you like to think about it.
Another technology which enables the creation of deepfakes is a GAN – Generative Adversial Networks – which is a kind of unsupervised machine learning.
How do you suss out a Deepfake?
This looks to be a game of never ending catch up with each. Because every time the good guys find a way of discovering deepfakes, the bad actors, may I say, fix the ‘vulnerability’ in their malicious code. For e.g. it was known that deepfake characters don’t blink their eyes, and this could have been because they were trained on images which didn’t have their eyes shut. The good guys got wind of it and used it but not for long. The bad actors got wind of it too and before you know it, we had deepfake characters who could blink.
There are industries putting in their might behind these quest of eliminating the scourge of deepfakes. An example is the Facebook driven ‘Deepfake Detection Challenge Dataset’ which is used to measure the progress on deepfake detection technology.
How rogue can it get?
Sadly, the applications of Deepfake has mostly always being malicious. After a while, it just sucks out the oxygen from the room to discuss or focus attention on the real important topics. From our side, all we can do is to treat any outrageous news from a unverified source with the right amount of skepticism