Do these attacks require fine control over the input? Eg. if you are scanning a patient’s body, does it matter if the model can be fooled by editing the values of individual pixels? This implies that you have a threat model where the data coming from the sensor is being manipulated, in which case all bets are off (the image could be entirely replaced). It doesn’t seem much different from a statistical model that you can blow up by feeding in values designed to cause a divide by zero (values that wouldn’t appear in real-world data).
It seems like a problem when classifying user-provided images (eg. identifying obscene images on a social network) but not so relevant when you own the sensors.