Challenges in leveraging vibration-based covert channels using machine learning techniques

Smartphones are increasingly equipped with sensitive accelerometers that can analyse acoustic vibrations on a physical surface. This allows them to gain a covert understanding of the surrounding environment by combining accelerometer sampling with sophisticated signal processing techniques. In this work, we analyse keyboard-sniffing attacks based on acoustic (vibration) covert channels, launched from a malicious application installed on a smartphone. An important requirement of such attacks is (a) access to reliable acoustic signals that can be distinguished from the noise floor by (b) applying appropriate signal processing techniques. Our analysis indicates that state-of-the-art attack techniques are fragile; injecting randomised noise via the vibration medium into the accelerometer, reduces the efficiency of the attack’s success rate from 80% down to the point of uselessness. We conclude that our work presents an intermediate step towards disabling the covert channel and ensuring full security.