F
Member
【Editorial Discussion】
A training report may advertise differential privacy while leaving “one contribution” ambiguous. If a person supplies many records, protecting the addition or removal of one record is not the same claim as protecting the addition or removal of that person's entire history. A privacy budget without its neighboring-dataset definition can invite a stronger interpretation than the mechanism supports.
This is not a question about deleting a saved memory after deployment. It concerns the unit of protection during private training and whether data preparation, sampling, clipping, and accounting all implement the same unit.
**Write two neighboring datasets**
In a synthetic example, create several records belonging to one fictional contributor. For a record-level claim, construct the neighbor by removing one record. For a user-level add/remove claim, remove all that contributor's records. Ask which pair the stated guarantee actually covers. Merely grouping records under a user identifier does not convert a record-private mechanism into a user-private one.
Record-level guarantees can sometimes imply weaker group guarantees under explicit contribution bounds and composition analysis. That is not permission to reuse the same budget unchanged for an arbitrarily prolific contributor. State the assumptions and the derived guarantee rather than allowing a user-level interpretation to arise by implication.
**Trace the contribution through one update**
Inspect a noiseless diagnostic of the pre-noise aggregation using only synthetic data. Increase the number of records assigned to the same contributor while keeping other contributors fixed. Does the contributor's total bounded influence follow the claimed construction, or can repeated individually clipped records accumulate beyond that bound?
One possible user-level construction aggregates a user's update and clips that aggregate before combining users. Other mechanisms are possible, but each needs its own sensitivity and accounting argument. A test that confirms clipping in one implementation cannot establish that the noise distribution or accountant is correct.
Check the sampler separately. Sampling records and sampling contributors are different procedures. If the accountant assumes a contributor-sampling scheme while the loader samples records, prolific contributors may participate under a different distribution from the one analyzed. Repeated appearances across steps and repeated training releases also belong in the accounting scope.
Finally, test preprocessing boundaries: can one person's activity be split into multiple identities, or can a contribution cap be bypassed by records arriving in separate files? The system needs an operational definition of contributor identity and a clearly stated limit where identity linkage is unavailable.
There are important limits. These diagnostics reveal mismatches; they do not prove differential privacy. Formal guarantees require the complete mechanism and assumptions, and empirical extraction tests cannot substitute for that analysis. Identity resolution can itself create privacy risks, so collecting more identifiers is not an automatic repair.
What evidence would you require before interpreting a record-level privacy report as protection for a whole contributor? How do you test that the loader, contribution bound, and accountant refer to the same unit?
A training report may advertise differential privacy while leaving “one contribution” ambiguous. If a person supplies many records, protecting the addition or removal of one record is not the same claim as protecting the addition or removal of that person's entire history. A privacy budget without its neighboring-dataset definition can invite a stronger interpretation than the mechanism supports.
This is not a question about deleting a saved memory after deployment. It concerns the unit of protection during private training and whether data preparation, sampling, clipping, and accounting all implement the same unit.
**Write two neighboring datasets**
In a synthetic example, create several records belonging to one fictional contributor. For a record-level claim, construct the neighbor by removing one record. For a user-level add/remove claim, remove all that contributor's records. Ask which pair the stated guarantee actually covers. Merely grouping records under a user identifier does not convert a record-private mechanism into a user-private one.
Record-level guarantees can sometimes imply weaker group guarantees under explicit contribution bounds and composition analysis. That is not permission to reuse the same budget unchanged for an arbitrarily prolific contributor. State the assumptions and the derived guarantee rather than allowing a user-level interpretation to arise by implication.
**Trace the contribution through one update**
Inspect a noiseless diagnostic of the pre-noise aggregation using only synthetic data. Increase the number of records assigned to the same contributor while keeping other contributors fixed. Does the contributor's total bounded influence follow the claimed construction, or can repeated individually clipped records accumulate beyond that bound?
One possible user-level construction aggregates a user's update and clips that aggregate before combining users. Other mechanisms are possible, but each needs its own sensitivity and accounting argument. A test that confirms clipping in one implementation cannot establish that the noise distribution or accountant is correct.
Check the sampler separately. Sampling records and sampling contributors are different procedures. If the accountant assumes a contributor-sampling scheme while the loader samples records, prolific contributors may participate under a different distribution from the one analyzed. Repeated appearances across steps and repeated training releases also belong in the accounting scope.
Finally, test preprocessing boundaries: can one person's activity be split into multiple identities, or can a contribution cap be bypassed by records arriving in separate files? The system needs an operational definition of contributor identity and a clearly stated limit where identity linkage is unavailable.
There are important limits. These diagnostics reveal mismatches; they do not prove differential privacy. Formal guarantees require the complete mechanism and assumptions, and empirical extraction tests cannot substitute for that analysis. Identity resolution can itself create privacy risks, so collecting more identifiers is not an automatic repair.
What evidence would you require before interpreting a record-level privacy report as protection for a whole contributor? How do you test that the loader, contribution bound, and accountant refer to the same unit?