Autenticação contínua usando sensores inerciais dos smartphones e aprendizagem profunda

Abstract

Many users have chosen to use mobile devices such as smartphones to perform day-to-day tasks such as sending emails, interacting with social networks, paying bills, and other banking transactions. Whilst such tasks have become simpler to perform, however, a large volume of sensitive and confidential information is stored and accessed from these devices, such as photos, bank logins and passwords, personal data, and electronic messages. When prioritizing the ease and usability of smartphones, the user can unknowingly neglect the security and privacy of sensitive data. To ensure the security of this data, most systems currently employ static authentication solutions. This is where the user unlocks the device once through an authentication mechanism such as password, grid pattern, security key, PIN (Personal Identification Number) or fingerprint sensor. However this security measure is vulnerable: in a scenario where an imposter user has access to passwords or gets physical access to the unlocked device, the entire amount of data will be exposed. To deal with this problem, this work proposes the development of a continuous authentication method for mobile devices using data from inertial sensors. The process of identifying the genuine or imposter user is performed through an authentication model defined from a deep network architecture based on convolutional neural networks with recurrent layers Long Short-Term Memory (LSTM). In addition, this work employs a trust model to avoid blocking genuine users and preventing an imposter from being undetected for a long time. Tests using data from 30 users show that the proposed model can detect imposter users in up to 61 seconds. These promising results prove the feasibility of using data from inertial sensors to define continuous authentication models.