Subliminal Probing for Private Information via EEG-Based BCI Devices
File under: “That’ll never happen, but has” via @cryptomeorg:
9 Conclusion In this work we have examined the question if subliminal attacks to users of EEG-based braincomputer interfaces (BCIs) are feasible. We have designed a proof-of-concept experiment in which the attacker tries to infer if the user knows a particular person or not, without the user noticing that she is being attacked. We hid visual stimuli in form of portrait photos of Barack Obama in a video as well as other visual stimuli that serve as a contrast. In an experiment with 27 subjects we find that our naive attack strategy is able to obtain 66% accuracy in predicting that a subject is familiar with Barack Obama, while an advanced attack strategy that incorporates confidence levels is able to improve the accuracy to 90%. The subjects achieved different levels of recognition in terms of detecting the manipulation of the video. At each recognition level, the attack was successful for most users including the users that did not notice any manipulation. 16 Our subliminal attacks have been carried out in a controlled setting to demonstrate their feasibility. Future research directions include exploration of different pathways for improving the attack, such as more sophisticated hiding mechanisms and internal subliminal validation techniques. The findings presented in this work suggest that BCI software with the full access to raw EEG data of users constitutes a new attack vector to user privacy and user secrets (arXiv:1312.6052v1 [cs.CR] 20 Dec 2013).
Written on December 23, 2013