How to recognize a deepfake?

Let’s face it – spotting deepfakes. i.e. an altered or AI-created video is only going to be easy in one case… when those who make them do shoddy work. But this happens quite often and we do have a chance to successfully spot deepfakes. 
 

By Piotr Henzler

Videos that present situations that never happened, people, events or utterances that never existed are plentiful. If we add manipulated images, we can safely assume that we are likely to come across a deep fake on a daily basis. But only occasionally it will be a perfect deep fake, one that cannot be spotted. Why? Because there are many people who will believe what they see even if a video is still a bit shoddy. People tend to accept without analysing, do not pay attention to details, they follow their emotions, because of social proof, because they want to make quick money, because they want… The fraudsters really do not need to be very meticulous. 
This presents an opportunity for consumers who do pay attention and can thus quite quickly recognize material that has been manipulated or completely made up. How? First, by following some basic rules of critical thinking and e.g. the CRAAP model presented in the text How to recognize fake news? Second (and third), by analysing deepfakes to detect typical errors and by using AI, which often creates such material, to apply the countermeasures, i.e. to detect deepfakes. The following recommendations focus on videos, but please note that many will also apply to photographs (or AI-generated images that look like photographs).  

Typical deepfake errors

While analysing videos, it is useful to take a closer look at all their elements, i.e.  the characters, their voice, their way of speaking, the background, the quality of the material, etc. Each of these may be a red flag, suggesting that we are dealing with a fake reality.

Who are the protagonists?
Have a look at the character in the video. Do you think it is an AI-generated avatar or a real person? If you suspect it may be an avatar, try to verify this. Do they look natural? For example, check how many fingers they have – is it not six or four? Is their skin tone even? Do their lip movement reflect the text they are saying? And when they move, do they leave a faint ‘trace’ where they used to be? If you see something suspicious, there is a risk that you are dealing with a fake. Though we should remember that people may, indeed, have different number of fingers or have skin discolorations…
If it is a real person, it may either be someone anonymous (‘an ordinary human’) or someone widely known (acting as an ‘authority figure’ in the message). If they are ‘ordinary’ people, you will not be able to verify if they are real. But if they are a scientist, a celebrity or a politician – you better have a closer look. Is the topic they discuss consistent with their social or professional role? A real life example – a ‘real’ bishop Kazimierz Nycz advertising a leg pain relief medication… Potentially plausible, but is it real? Or – Elon Musk speaking Polish and encouraging people to invest in… He may have mastered our language to do the commercial, but is it true?

What do you hear?
Please note the way in which the character speaks. Does their voice sound natural, in line with their age or gender? Is the intonation correct? Do they use word in a foreign language or have incongruent pronunciation? 
Pay attention to the body language. Is their voice quiet and calm while they gesture a lot? Or are they speaking about serious or sad matters in a voice that is joyful and energetic? This may suggest that the voice comes from a very different footage.

What do you see?
We have already talked about lip-sync, but now please have another look at the speaker’s lips. Is the image somehow blurred as they speak? Or has their face shape changed unnaturally? Have a look at their teeth. Deepfake apps sometimes ‘drop’ teeth, either all or some, as they generate manipulated footage.
It is also worthwhile to look at the eyes. Is the gaze natural and alive or blank and motionless? Does the person blink (or not) and if they do – do they blink with natural frequency?
If they are wearing earrings, pendants or eyeglasses, do these look natural and keep their shape, is there a faint trace when there is movement?
Also, look at the larger frame, behind the speaker. What is in the background? Is the background congruent with the topic or the person? Are there no strange, out-of-place objects? Is the background undulating, is it not dissolving?

AI to the rescue

Should such a technical analysis bring no conclusive answer to the question if we are indeed dealing with a deepfake or not, we may ask AI for help. Yes, it does contribute to deepfake generation, but it can still be useful in verifying the veracity of videos (or images). Below, you will read about some such apps, but there are many more.   
A suspicious material can be uploaded (as a link or a file) to e.g. Scanner.deepware.ai for assessment – if it is real or fake. Somewhat simpler but still good tools are provided by Attestiv.video or Deeepfake Detector, which can also assess pictures. A good picture analysis (and also text-analysis, whether it is AI-generated or not) is provided by Isitai.com. The majority of such service providers (these are but few examples) allow initial free-of-charge analysis of a few such images, but then charge fees. You may also consider checking the material by two separate providers, to make the results more viable.
If this is still not enough, there is yet another fact-checking phase you can consider – checking who is behind the disseminated message (e.g. is the social media profile that disseminated the material a real person or a bot), who owns the website where the material was first published, and is the IP of the website where the video originates trustworthy or not. How to do it? Just enter a relevant question in your web browser and search for appropriate apps that can do it for you. Good luck!