Latent Representation Learning and Manipulation for Privacy-Preserving Sensor Data Analytics

ObscureNet

“…These solutions are usually tailored to specific use cases and are incapable of preventing the leak of private information while maintaining the utility of data. The latter category includes federated learning frameworks [28], differential privacy algorithms [7,31], cryptographic solutions based on homomorphic encryption and compressive sensing [1,34], and privacy-preserving techniques that transform data to a subspace where private attributes can be easily identified and altered [8,11,24]. Federated learning addresses the problem of training a model given data from many users without transferring them to a server.…”

“…This approach provides a reasonable trade-off between utility and privacy by minimizing the leak of private information while preserving the information content of the input data. However, the data anonymized by these networks is shown to be susceptible to the re-identification attack [11]. To address this shortcoming, a probabilistic transformation technique is proposed in [11] which manipulates private attributes such that the anonymized data is less vulnerable to the re-identification attack.…”

“…However, the data anonymized by these networks is shown to be susceptible to the re-identification attack [11]. To address this shortcoming, a probabilistic transformation technique is proposed in [11] which manipulates private attributes such that the anonymized data is less vulnerable to the re-identification attack. This approach has several disadvantages.…”

“…In recent years, deep generative models, and in particular, encoderdecoder architectures have been used to manipulate natural images, e.g., to change the facial expression of a portrait [18,19]. Inspired by the success of autoencoders and generative adversarial networks in the image generation task, they have been applied to anonymize data generated by IoT and mobile devices [11,24,25]. These approaches transform sensor data, such that the leak of private information is minimized.…”

“…In follow-on work [24], an autoencoder is trained in an adversarial fashion using a loss function that ensures maximum data distortion and minimum mutual information between private attributes and synthesized data. However, the sensor data anonymized via both techniques can still be used to re-identify the user and their private information should the attacker pass a dataset with known private attributes through these autoencoders [11]. We refer to this as the re-identification attack.…”