Browsing by Author "Aguiar, Armani"

Brain-computer interfaces (BCIs) are systems that measure and interpret electrical signals from the brain’s nervous system. One of the most common ways of measuring these brain signals is through a non-invasive technique called electroencephalography (EEG), which places electrodes on the subject’s scalp. There are generally three components: the physical hardware that acquires the signals, the software which does pre-processing and machine learning (through feature extraction/classification), and a communication channel that sends the classification to a control system to produce commands. This paper focuses on the software component by looking at the machine-learning models used and their vulnerability to data poisoning and evasion attacks, as well as proposing some solutions to these attacks. These solutions include authenticating data using unique features found in EEG signals, a ‘denoiser’ model trained from generated adversarial examples, and an RFID sensor placed within the BCI that authenticates the data with an external reader. Because this interface will likely see more adoption in healthcare, the objective of this paper is to bring its problems to surface and inform researchers how existing designs can be improved upon to protect sensitive brain data and the quality of life for disabled patients.