Polynomial-time Attack on Output Perturbation Sanitizers for Real-valued Databases
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
We review the attack given by Dinur and Nissim [6] on the output perturbation sanitizer, and generalize it to a setting that includes, as particular cases, databases with values in {0,1}---with the metric considered in [6]---and databases with real values, with other appropriate metrics (hence the binary case is not included in the real case). Previous works [12, 14] on the binary case gave results more efficient than ours. Those results could be used to extend the binary case to the real-valued case, hence implying our results. The contributions of this paper are: to make the implication explicit, and to give an alternative general proof. We state a property about the function dist that measures the error of the attacker's approximation of the database, which is satisfied in our cases of interest, and is sufficiently strong to prove the impossibility results regarding the privacy provided by the output-perturbation sanitizer, in both the real and binary cases. In this general context we establish an inequality (an upper bound to the probability of adversary's failure) that relates all the parameters of the problem---the size of the database, the relative error of the adversary, the number of queries made by the adversary (which determines its time complexity), its probability of failure, and the perturbation of the sanitizer---making explicit the trade-offs among them. From this inequality we deduce that for binary and real valued databases, the adversary described in [6] can defeat perturbation o(n1/2) with time complexity determined by o(n log n) number of queries (instead of O(n log2 n) as in [6]).
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.001 |
| Open science | 0.001 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it