Selective EEG Signal Anonymization using Multi-Objective Autoencoders

The availability of low-cost brain-computer interfaces and related software has enabled application developers to access this technology seamlessly. However, unsupervised access to users' brain signals raises alarms for EEG data privacy and identity protection. This paper explores a new direction to selectively anonymize a person's brain signals resulting from a response to a stimulus. These time-locked potentials containing sensitive user information are masked to allow for the intended task/event classification (brain task activity) while minimizing the accuracy of subject classification (brain identity activity).We study the feasibility of an autoencoder architecture, enveloped with regularizers and a multi-objective loss function, to achieve an optimal utility-privacy trade-off for EEG data application. We observed a drop of 35% in the accuracy of the subject's classification, while suffering only a loss of 14% in the accuracy of the task classification. This algorithm can be applied to multi-channel and multi-subject scenarios, and our results demonstrate a proof-of-concept that we can generalize an anonymizing autoencoder architecture to be applicable to intricate stochastic data such as EEG.