De-identifying transmission system using wireless channel as differential privacy noise and deep neural networks

Abstract

In order to deal with the unprecedented problem of data privacy in deep-learning based computer vision technologies, we propose a de-identifying transmission system that utilizes a wireless channel as differential privacy (DP) noise and is constructed using deep neural networks. By following the Gaussian mechanism of DP, we present that the signal received at a receiver can be considered as a Gaussian mechanism. Then, we discuss the relationship between the signal-to-noise ratio of a received signal and the privacy budget of DP, and introduce how to control transmit power to achieve a specific privacy budget. The proposed system can be divided into three parts, named transmitter, wireless channel, and receiver. The transmitter and receiver are constructed by using deep-learning networks. The transmitter consists of an image encoder network based on a neural network and a power control module that controls the transmit power for adjusting the level of de-identification. The wireless channel acts as differential privacy noise, and anonymizes the transmitted image feature vector extracted from the transmitter’s image encoder network. The receiver includes a post-processing network and an image decoder network, which are also implemented by using deep neural networks. The post-processing network is proposed for high-quality decoded face images at the receiver, which maps received feature vectors perturbed by a wireless channel back into the latent space of deep-learning based image encoder and decoder networks. Finally, extensive qualitative and quantitative evaluations confirm that the proposed system can well de-identify transmitted face images only by controlling the transmit power while maintaining the usefulness of decoded face images. The proposed system shows that the Recall@1 is smaller than 5.2 and the face detection probability is larger than 90 % at SNR=2 dB. Since the de-identification process is performed by wireless channel noise, the proposed system does not require de-identification processing at a transmitter, and thus there is no burden on the transmitter to anonymize face images. Moreover, the additional advantage of this proposed system is that the level of de-identification can be controlled only by changing the transmit power. This proposed de-identification system can be utilized in a wireless image acquisition scenario where images are captured from a wireless edge, such as a CCTV camera, and then sent to a server while protecting people’s privacy.