Learning Semi-Supervised Anonymized Representations by Mutual Information

Abstract

This paper addresses the problem of removing from a set of data (here images) a given private information, while still allowing other utilities on the processed data. This is obtained by training concurrently a GAN-like discriminator and an autoencoder. The optimization of the resulting structure involves a novel surrogate of the misclassification probability of the information to remove. Several examples are given, demonstrating that a good level of privacy can be obtained on images at the cost of the introduction of very small artifacts.

Subverting Privacy-Preserving GANs: Hiding Secrets in Sanitized Images

“…Hence, an empirical, task-dependent privacy checks are used to provide a holistic measure of privacy. Recent work leverages adversarial networks to sanitize input images and adopts similar DL-based discriminators for privacy examination (Edwards and Storkey 2016;Raval, Machanavajjhala, and Cox 2017;Pittaluga, Koppal, and Chakrabarti 2019;Chen, Konrad, and Ishwar 2018;Wu et al 2018;Tseng and Wu 2020;Feutry, Piantanida, and Duhamel 2020;Maximov, Elezi, and Leal-Taixé 2020;Xiong et al 2019). These works employs adversarial training to jointly optimize both privacy and utility objectives.…”

“…These so-called "privacy-preserving GANs" (PP-GANs) can sanitize images of human faces such that only their facial expressions are preserved while other identifying information is replaced (Chen, Konrad, and Ishwar 2018). Other examples include: removing location-relevant information from vehicular camera data (Xiong et al 2019), obfuscating the identify of the person who produced a handwriting sample (Feutry, Piantanida, and Duhamel 2020), and removal of barcodes from images (Raval, Machanavajjhala, and Cox 2017). Given the expertise required to train such models, one expects that users will need to acquire a privacy preservation tool from a third party or outsource GAN training, so proper privacy evaluation is paramount.…”

“…The privacy offered by PP-GANs is typically measured using empirical metrics of information leakage (Chen, Konrad, and Ishwar 2018;Xiong et al 2019;Feutry, Piantanida, and Duhamel 2020). For instance, Chen, Konrad, and Ish-war (2018) use the (in)ability of deep learning (DL)-based discriminators to identify secret information from sanitized images as the metric for privacy protection.…”

Generating Synthetic ECGs Using GANs for Anonymizing Healthcare Data

“…In this context, anonymization arises as a tool to mitigate the risks of obtaining and massively processing personal data [4]. We propose GAN-based anonymization [5] of private health data, so a seedbed would be obtained from the training data that allow not only to capture information from the original data, but to generate new information with a similar behaviour.…”

Subverting Privacy-Preserving GANs: Hiding Secrets in Sanitized Images

“…These so-called "privacy-preserving GANs" (PP-GANs) can sanitize images of human faces such that only their facial expressions are preserved while other identifying information is replaced (Chen, Konrad, and Ishwar 2018). Other examples include: removing location-relevant information from vehicular camera data (Xiong et al 2019), obfuscating the identity of the person who produced a handwriting sample (Feutry, Piantanida, and Duhamel 2020), and removal of barcodes from images (Raval, Machanavajjhala, and Cox 2017). Given the expertise required to train such models, one expects that users will need to acquire a privacy preservation tool from a third party or outsource GAN training, so proper privacy evaluation is paramount.…”