Face De-identification with perfect privacy protection

Abstract

The rising concern for privacy protection and the associated legal and social responsibilities have led to extensive research into the field of face de-identification over the last decade. To date, the most successful algorithms developed for face de-identification are those based on the k-Same deidentification, which guarantee a recognition rate lower than 1/k. However, the current k-Same solutions such as k-Same-Eigen and k-Same-M all rely on a decent value of k to deliver a good privacy protection. This paper proposes a departure from a fundamental aspect shared by the current k-Same solutions and thereby introduces a new member to the family which achieves perfect privacy protection for any original face regardless of the value of k.

Retaining expressions on de-identified faces

“…The recognition rates of k-Same-M-Select faces stay slightly below the theoretical maximum of 1/݇. As proved and tested in [9], k-Same-furthest without the FET process produces a recognition rate of zero regardless of the value of k. The recognition rates of the k-Same-furthest-FET faces is nearly zero regardless of k, indicating that the expression transfer process after k-Same-furthest de-identification has hardly any impact on the privacy protection performance.…”

“…The newly introduced member of the k-Same family, kSame-furthest [9], is a model-based approach, representing faces as AAM features to avoid ghosting artefacts. It identifies the k faces that are furthest away from a given probe face image, calculates the average of these faces, and returns this average as the de-identified face of the probe.…”

“…It identifies the k faces that are furthest away from a given probe face image, calculates the average of these faces, and returns this average as the de-identified face of the probe. The k-Samefurthest algorithm guarantees that face de-identified by it can never be recognised as the original face as long as the identity distance measure used to recognise the de-identified faces is the same as that used by k-Same-furthest [9]. If the deidentified faces are recognised by Eigenface technique [16] which represents faces in their PCA space rather than AAM, kSame-furthest offers a recognition rate very close to zero [9].…”

“…To remove ghosting artefacts, the k-Same-M algorithm [7] models each of the k closest images as an Active Appearance Model (AAM) [8] and then performs averaging on the AAM model parameter vectors. To truly protect privacy, a new member named k-Same-furthest has been introduced to the k-Same family [9]. Like k-Same-M, k-Same-furthest is model based and represents face images as AAMs.…”

De-identification for privacy protection in multimedia content: A survey

“…For both algorithms, there are two main problems: they operate on a closed set I, and the determination of the proper privacy constraint k. In order to produce de-identified images of much better quality and preserve the data utility, the Model-based k-Same algorithms [70] are proposed -one of which is based on Active Appearance Models (AAMs) [72] and another based on the model that is the result of mixtures of identity and non-identity components obtained by factorizing the input images. Modifications to the k-Same Select algorithm, in order to improve the naturalness of the deidentified face images (by retaining face expression) and privacy protection, are proposed in [73,74].…”