| dc.description.abstract | This thesis aims to investigate how Apple's NeuralHash algorithm can be used in the context of facial recognition to improve privacy in facial recognition systems. Existing facial recognition solutions rely on having facial images available to match identities, however, this can impair the privacy of individuals, as the images can contain sensitive information that the individuals do not want to share. In this thesis, the NeuralHash algorithm is used to hash facial images of subjects in the ColorFERET Dataset, and the NeuralHashes are compared to attempt to identify the same subjects and different subjects. The NeuralHash algorithm's ability to hide information is also investigated, in addition to collision- and evasion attacks on NeuralHash. The results show that using a threshold of approximately 0.24, the false acceptance rate and false rejection rate are 9.68 %. If the threshold is set to 0.1, the false acceptance rate drops to 0.16 %, while the false rejection rate rises to 31.45 %. Some general information about images such as gender can be inferred from the NeuralHash, while more nuanced information is not retrievable. Gradient-based attacks can be used against NeuralHash to evade collisions, and to force collisions with a target NeuralHash. | |
