In Section 4.1 of the paper, you mention:

We quantify the ability to anonymize images by measuring the “re-identification rate” (defined as the number of images whose identity is still detected in the anonymized version, over the total number of images) using FaceNet.

Can you please confirm exactly how the above is done? I assume you pass in both the original and anonymised face image through the FaceNet model to get the respective embeddings. Do you then just calculate cosine similarity between the embeddings of original and anonymised face, and if similarity is higher than a threshold the image is considered "re-identifiable"?

If the above is exactly what is being done, then what threshold did you set in your experiments in the paper?

Thanks,
Rohit