In the final thesis, common deidentification methods are presented with the
goal of concealing identity. The aim is to design a deidentification process
through image preprocessing and the use of existing concepts of anonymity,
depth images, and latent diffusion models. The paper first reviews the state
of the art and presents several related approaches. It then introduces the concepts
used in our implementation, the image datasets employed, and the evaluation
methods. Using the AUC metric, where our best result reaches a value
of 0.75, we demonstrate the success of deidentification. We also compare the
preservation of other attributes, such as facial expression, gender, and ethnicity.
Additionally, we evaluate results using the Equal Error Rate (EER),
where our model achieves competitive values compared to established methods
such as LDFA and FAMS. Across different datasets—including RaFD,
CelebA-HQ, and XM2VTS—our model achieves EER scores ranging from
19.91% to 31.28%, demonstrating a strong balance between identity protection
and image utility. These results confirm that our approach effectively
reduces recognizability while preserving image quality.
|
