It's highly unlikely they obtained 17 billion samples, so they're likely guesstimating it by assuming each attribute is independent, and summing the entropy of all attributes. That's obviously incorrect, both because attributes are inevitably going to be correlated (eg. ip geolocation correlated with time zone), and that two identical devices (eg. 2 iPhones) will have identical fingerprints.