Privacy Issues Regarding the Application of DNNs to Activity-Recognition using Wearables and Its Countermeasures by Use of Adversarial Training

Abstract

Deep neural networks have been successfully applied to activity recognition with wearables in terms of recognition performance. However, the black-box nature of neural networks could lead to privacy concerns. Namely, generally it is hard to expect what neural networks learn from data, and so they possibly learn features that highly discriminate user-information unintentionally, which increases the risk of information-disclosure. In this study, we analyzed the features learned by conventional deep neural networks when applied to data of wearables to confirm this phenomenon. Based on the results of our analysis, we propose the use of an adversarial training framework to suppress the risk of sensitive/unintended information disclosure. Our proposed model considers both an adversarial user classifier and a regular activity-classifier during training, which allows the model to learn representations that help the classifier to distinguish the activities but which, at the same time, prevents it from accessing user-discriminative information. This paper provides an empirical validation of the privacy issue and efficacy of the proposed method using three activity recognition tasks based on data of wearables. The empirical validation shows that our proposed method suppresses the concerns without any significant performance degradation, compared to conventional deep nets on all three tasks.

RETRACTED ARTICLE: Human activity recognition with deep learning: overview, challenges and possibilities

“…In specific, its black-box characteristics may inadvertently expose user-discriminatory characteristics for the deep learning technique. The authors investigated the problem of privacy using CNN technology for detection of human behavior in [143]. The empirical studies indicate that while CNN is qualified only for behavior detection with a lack of cross-entropy, the learned CNN characteristics have also shown powerful abilities for consumer discrimination.…”

“…To fix this issue, some researchers have investigated the use of an enemy failure feature to minimize the discrediting quality of the data during the training phase. For example, in order to minimize the user identification precision, Iwasawa et al [143] proposed adding an adverse outcomes failure into the normal operation description missed. The developers of [144] and [145] have also taken the same notion to avoid loss of information.…”

Preventing Sensitive Information Leakage From Mobile Sensor Signals via Integrative Transformation

“…Our lives become safer and more convenient with the assist of these personalized and ubiquitous services. Nevertheless, the concern of privacy leakage of this kind of personal data has drawn an increasing attention in recent years [3]. Although the sensors are originally used to capture the movement of users, personal traits can also be hold within the continuous signals unintentionally.…”

“…Therefore, the subject's weight information could be determined through interpreting the sensor signals. Other personal information like age, gender, and height can also be figured out in the similar way [3]- [5]. This kind of information leakage is unacceptable so that the sensor data cannot be directly sent out without any privacy destruction procedure.…”

“…The first solution is non-intelligent that are only damaging to the following applications yet not able to fully prevent information leakage at all time [8], [9]. In contrast, the latter solution is a more practical and promising approach that engages boosting attentions [3], [10], [11].…”