# Processing with AI
## Exploration : IA & Ethics

Name:
> Mathieu Charret
>
Subject:
> Detect student presence in class using Face Recognition

>[TOC]

## Design brief

### Biais

If we don't source our dataset with enough rigor, the following biais might appear:

>1. Some people might not be recognized because of their skin color, the brightness in the room, the quality of the camera, the period of the year (tanned skins), face injuries etc...
>2. People could be recognized as someone else from the class.
>3. Some people might not be detected at all. The computer would consider that no one is in front of the camera.

We will ensure that our model is not biaised by:

>1. Sourcing our data from pictures of various skin colors. Representing all races and origins.
>
>2. Pictures in various environment will be used as data in order to avoid problems due to :
>a. Quality of camera
>b. Brightness in the room
>c.
>
>3. Students will have to take a reference pictures every 2 months with the camera they use for the classes.
3 pictures :
>a. One facing the camera
>b. One showing their right side
>c. One showing their left side
>
>The reference pictures will be compared to the pictures on their student cards by humans.
>
>In case of face injuries or detection troubles, students will be able to retake the reference pictures manually.

### Overfitting

We will make sure our model does not overfit by
>1. Checking the accuracy of our model on a sample of the students, representing all origins, in different environment with different webcam quality.
>2. Sourcing or data from an equal amount of pictures from different origins, with different environments and webcam quality.

### Misuse

>We have to remind ourselves that our application could be misused by **students** to **steal pictures from one another**. Moreover, our application could be fooled by some pieces of software that **simulate the use of a camera and broadcast a loop video.**

### Data leakage
*Choose the most relevant proposition:*

>In a catastrophic scenario, where all of our training dataset were stolen or recovered from our model, the risk would be to get sued by students that gave access to their reference pictures to the school only.
>
>One way to limit the risk is to delete the old reference pictures every time we take new ones. We should also encrypt our database and store it in European servers in order to respect the RGPD laws.

### Hacking

> If someone found a way to "cheat" our model and make it make any prediction that it want instead of the real one, the risk would be that this person could sell his knowledge to the other students for money.
>
> If 20% of the student hear about this tip, our application would be useless as we couldn't differentiate present students from the others. Resulting in a status quo.