Consider Alice's census response as a row in a table of responses. You can reconstruct every attribute/column for which a sufficiently large number of summary statistics are released.
Names are not a part of the statistics the Census releases, so you won't be able to reconstruct the name. Make a fingerprint out of some of the reconstructed attributes and run a database join against another dataset with names. You've now enriched your data with any remaining attributes that were not a part of your fingerprint.
I'm grateful no one has attacked and shared the 2010 decennial census, or ACS, which has considerably more questions. If it seems far-fetched... well, you only need one person to do it, there's an existence proof, and the attack is basically just a convex optimization problem.
Extra:
I've written some reconstruction attacks like this myself. One approach is to find the least squares solution to |Ax - b|. Let b be a vector of Census statistics. A is a query matrix. An attacker has both. Solve for x, which is a column in the dataset. If b is long enough, then the system is fully determined, and you can solve for x exactly. In practice, b can be much smaller than x for the system to reconstruct x with high accuracy. Repeat for each column. There are more efficient approaches, like SAT solvers, for large systems.
Happy to talk more about the general approaches used to privatize census data if anyone is curious. I don't work on the Census mechanisms, but I do privacy research and am familiar with DP hierarchical histogram mechanisms.
It would be possible to write a learned program to simulate browsing habits. It is ironic, in a way, that you would need to track yourself, to avoid tracking.
I use Firefox personally, but develop with Chrome. I have a hard time getting stack traces from Firefox reliably. Oftentimes one of the following occurs on Firefox:
There is no error, but the page won't load.
There is an error, but the error message is generic.
There is an error message, but no stack trace.
When I switch to chrome, I always get the error and stack trace without any further code modifications.
I make sure to clean up for Firefox in releases. But it's just impossible to develop without error messages and traces. I'm on Ubuntu using the latest stable releases.
I think his concern is- since the security program would need to run on the attacker's computer, the attacker may simply opt not to use it, circumvent it. That being said, I don't see why it can't be used to sanitize requests to the server.
Names are not a part of the statistics the Census releases, so you won't be able to reconstruct the name. Make a fingerprint out of some of the reconstructed attributes and run a database join against another dataset with names. You've now enriched your data with any remaining attributes that were not a part of your fingerprint.
I'm grateful no one has attacked and shared the 2010 decennial census, or ACS, which has considerably more questions. If it seems far-fetched... well, you only need one person to do it, there's an existence proof, and the attack is basically just a convex optimization problem.
As to what can specifically be in those columns, it's pretty narrow for the decennial, but the ACS is more broad. Check out the ACS summary data: https://data.census.gov/cedsci/table?q=United%20States
Extra: I've written some reconstruction attacks like this myself. One approach is to find the least squares solution to |Ax - b|. Let b be a vector of Census statistics. A is a query matrix. An attacker has both. Solve for x, which is a column in the dataset. If b is long enough, then the system is fully determined, and you can solve for x exactly. In practice, b can be much smaller than x for the system to reconstruct x with high accuracy. Repeat for each column. There are more efficient approaches, like SAT solvers, for large systems.
Happy to talk more about the general approaches used to privatize census data if anyone is curious. I don't work on the Census mechanisms, but I do privacy research and am familiar with DP hierarchical histogram mechanisms.
See also Abowd's summary of the attack: https://blogs.cornell.edu/abowd/special-materials/245-2/