Has the Social Robot Been Hacked? An Emotion Detection System with Distance-Range Parameters for Feature Selection

As robots become more social and present in living spaces, they are also more likely to attract the attention of cyber attackers. Ranging in the level and method of communication with the users, social robots are able to collect, share and analyze a great variety and large volume of personal data. Adding to the danger of a hacked robot to be perceived as asocial or antisocial, it may also expose its users’ security and privacy. In this paper, we highlight important and unique intersections of social robots and cybersecurity. We also model one such interaction in order to explain the validity of our claim which is, disturbing impact on user emotion through the interaction with social robots. We design an emotion detection system for users that interact with social robots on a regular basis. The negative emotions that arise with the interaction with the robot could signal a misconfigured/inaccurately responding or compromised social robot. Our model is designed as a part of a future anomaly detection system, which should run on an external device that resides in the same space as the human and the robot to monitor the audio utterances. The emotion detection model has to be lightweight to install in simpler devices, fast and accurate in order to identify emotions in real time. Therefore we have designed a new parameter-based metric learning wrapper method for optimized feature selection in machine learning model, which generates fast and accurate results for binary classification of emotions, This also addresses the imbalanced class problem in the data. Our results on an acted, multi-speaker input set show comparable or better performance than state of the art methods.