Hands-On Artificial Intelligence for Cybersecurity

Facial recognition vulnerability to adversarial attacks

The reason why common facial-recognition models are vulnerable to adversarial attacks is that two identical CNNs are used, which together constitute a Siamese network. In an attempt to calculate the distance between two representative images of the faces to be compared, a CNN is combined with the first image and another CNN is combined with the second image.

The distance calculated between the representations—also known as output embeddings, formulated by the CNNs in relation to the respective images—is evaluated based on the exceedance of a given threshold value.

The weak link of this facial recognition method is constituted precisely by a correct evaluation of the distance existing ...