SonarID: Using Sonar to Identify Fingers on a Smartwatch

Single impulse response estimation representing current sonar reflections (1024 samples) Sonar Fingerprint Retains nSamples around peak Upsamples ZC Sequences by 2 Concatenates data from window of nSeqs consecutive ZC sequences CNN Classifier Input Image: nSamples by (nSeqs × 2) Finger idnetification Zadoff-Chu (ZC) Sequence (1024 samples) Hand reflection path Body reflection path Conv 2D (32) Conv 2D (32) Maxpooling Conv 2D (64) Maxpooling Conv 2D (128) Maxpooling Flatten Dense (1024, ReLU) Dense (128, ReLU) Dense (3, Softmax) Samples Sequences Smartwatch Audio Playback and Recording

Figure 1: Overview of SonarID: during a screen touch by the thumb, index, or middle fnger, a speaker on one side of a smartwatch emits an ultrasonic sonar signal (a Zadof-Chu (ZC) sequence, modulated over a carrier wave) and a microphone on the other side receives it. The signal is demodulated and processed to create a sonar fngerprint: a time-varying image, composed of nSeqs ZC sequences, each trimmed to nSamples in length, of the impulse response to the signal during the touch. A deep learning model processes this data to identify which fnger performed the touch.