Machine Learning Voice Synthesis for Intention Electromagnetic Interference Injection in Smart Speaker Devices

This work presents the effectiveness of using machine learning (ML) synthesized voice samples to control smart speaker devices through radiated intentional electromagnetic interference (I-EMI). In previous works, the feasibility of using I-EMI to control smart speaker devices was shown. However, devices that are trained to only recognize a single person’s voice or only execute certain commands from that person will not be as susceptible to this attack. By training a generative adversarial network (GAN) using samples of the target’s voice, this security feature can be bypassed directly, increasing the feasibility of the attack.