We can first begin with the discussion: What is a Deep Fake?
As technology Is advancing, so are the techniques to create malicious content that could be harmful to you as well. In standard terms, a deepfake is an image, audio, or Video created with technology that imitates a person’s appearance or voice.
There are a lot of technologies available now to create a deepfake. First of all, the physical movements and voices are analysed using advanced AI Technologies. Then the pictures or voice of a person is used to create a hyperrealistic image or audio using generative adversarial networks (GANs), appearing that some person actually did or said so. These deepfakes can be used to manipulate people using highly influential persons, deepfakes, false statements by any leader, or misleading people.
How to Identify a Deep Fake
Although it is very difficult to identify deepfakes with advanced technologies, we can still follow up with some methods to see if something in front of our eyes is natural or AI-based. To determine a deepfake, we can look for the following:
- FACIAL EXPRESSION: Identifying deep fakes starts with facial expression. We can check whether the expression, eye movement, or smile matches the word since deep fakes sometimes struggle with natural facial expressions. They may show unnatural smiles and eye movements. Deepfakes may show rigidity in facial expressions, exaggerated facial expressions or too frequent blinking.
- SKIN TONE AND TEXTURE: while in a deep fake, a face Is attached to the technology, there may be some unevenness in the skin tone of the face and the body. The face colour may not match the hand, and while making any deepfake video, the AI struggles with making the tone even. Hence, you can see uneven skin tone around the eye and neck.
- LIGHTENING OR INCONSISTENT BACKGROUND: The light in the background may not match the light on the face; this can be noticed in shadows as well. If some objects nearby can reflect, reflection can also be checked, as AI sometimes does a poor job on reflections and shadows while doing Deepfakes. In the case of a deepfake, there may be some distortion in the background, and the background may also wrap around the object. In some cases, if the subject moves through the area, the background can show unnatural movement compared to the subject.
- IRREGULARITY IN AUDIO: In the case of a deepfake, the audio may sound unnatural with an irregular tone, there may be a lack of emotions in the voice, and it may contain artificial noises in the background. If the Video is deep fake, that will lack any natural tone that comes along with any natural conversation with a human. The audio may also not go with the expression or gestures made by the person in the Video.
Until now, we have not seen how to analyse any Deepfake manually, but there are also AI tools that can detect deepfakes. Some are Google Reverse Image Search (for images), InVID (for Video), Deep ware Scanner, Microsoft Video Authenticator, etc. let us see how can we detect Deepfake using these AI tools:
As per above instruction there are levels of Deepfake detection through AI as well. We have many tools available for Deepfake detection that includes different kinds tools. We can check on them one by one:
Microsoft Video Authenticator: Helps in detecting deepfake by analysing any change in pixels, light or background in the video.
Deepware Scanner: it helps in scanning deepfake video through its program.
Sensity AI: Helps in detecting real-time deepfakes.
MIT Detect Fakes: It is a research program that focuses in detecting deepfakes through video analysing. One more Research tool to help detect Deepfake is FaceForensics++
The AI tools mentioned here scan for a video and give a detailed analyses on whether it is AI edited or not. The key features of this scanning include:
Facial Expression and Inconsistencies:
First of all the mistake made by an ai is in the eye blinking. The AI detects any unnatural eye movement or gazing direction. If the person does not maintain natural eye movement than the video may be a deepfake. Also previous models were not so good with eye movements but current models have improved and so their detection. The other part of facial expression that is lip syncing and facial symmetry is also traced by these detection tools and if there are any desynchronizations in the audio from video, it may be concluded as a deepfake.
Light and Shadow Analysis:
In natural videos, the shadow is aligned with the light source but in case of a deepfake it may show inconsistency and AI tools are capable of detecting reflections in the eye of the subject. If the image does not match the surrounding or something does not match with the real-world, AI detection will show them.
Pixel Level Analysis:
AI can detect any blur or unsmooth part that can occur around faces in deepfakes. AI can analyse frame rate of a video that is a sign of deepfake video.
As new tools are coming to create more realistic deepfakes, so are its solutions. AI detection tools are advancing as well that are able to detect deepfakes through pattern analysis, motion detection or forensic techniques.
