Even if I give you the ability to go from picture to real name (which to many is still fairly jarring) given only "open" databases at 100% accuracy you still can't come near your 50% mark. The leap from real name and place of birth to SSN is the limiting factor, and the reason for this is the randomness inherent in the assignment of the last 4 digits of one's social security number. Given the ability to guess > 10% of that range (they have access to all SSNs who are deceased as well), the researchers could still get only 8% nationwide accuracy on recent SSNs.
You would need a fundamental change in that algorithm to get close to the numbers you want. This isn't a technological limitation, it's a algorithmic limitation, and you can't really put time frames on these kinds of breakthroughs.
Current techniques that try to deal with such variation need several (clear) pictures from different angles to build a model of a face to be able to recognize a face.
I'd go with decades rather than years before the limitations mentioned are resolved.