Face Recognition systems: performance evaluation and bias analysis
| Type of publication: | Idiap-Com |
| Citation: | Dayer_Idiap-Com-04-2020 |
| Number: | Idiap-Com-04-2020 |
| Year: | 2020 |
| Month: | 8 |
| Institution: | Idiap |
| Address: | Rue Marconi 19, 1920 Martigny |
| Abstract: | User authentication is a crucial part of data security, and biometrics is an advantageous way of achieving this. Face images capture being minimally invasive and easy to acquire makes face recognition a good contender for being used in a lot of applications that require to know if the user is really who he claims he is. In this thesis, I compare the performance of multiple existing face recognition systems on different datasets. I then present how a convolutional neural network system works, and show the performance results of such a system trained from scratch for face recognition. I show that training a big neural network with few images is detrimental, and a big training dataset is required. An experiment on racial bias evaluation is then presented with methods to reduce the disparity between ethnicity in the products of a face recognition system. |
| Keywords: | Artificial intelligence, bias, Convolutional neural network, Face Recognition, neural network |
| Projects |
Idiap |
| Authors | |
| Added by: | [ADM] |
| Total mark: | 3 |
|
Attachments
|
|
|
|
|
Notes
|
|
|
|
|
