Anonymizing Face Images by Using Similarity-Based Metric

Vast numbers of face images are posted and circulated daily on social network and photo-sharing sites. Some face images are linked to the person's name, like those on user profile pages, while others are anonymized due to privacy concerns. If an anonymized face image is linked to a named one, that person's privacy is infringed. One way to overcome this privacy problem is to anonymize face images when they are posted on social networks. However, current face anonymization methods fail to meet two key requirements: being provably secure against de-anonymization and enabling users to control the trade-off between security and usability (similarity to the original face) of the anonymized face images. We are developing a similarity-based method for face anonymization that meets both requirements in those cases where a new face image of a person is to be posted when many face images including those of that person are already posted. The basic idea is to hide the new face image in s face images that are equally similar to the face image of the same person. We theoretically demonstrated that the probability of an attacker correctly linking the anonymized face image to an image of the same person is less than 1/s. We also showed theoretically and confirmed experimentally, with 150 sample face images, that the larger the s, the less usable the anonymized face image. The security of our method holds in spite of future improvements in face recognition tools.