Certified Attribute Privacy in Gan Latent Space
Bibliographic record
Abstract
In this paper, we address the challenge of selectively protecting specific semantic attributes in images while preserving the utility of others. We show that, under our mechanism, the released attribute scores satisfy$(\epsilon, \delta)$differential privacy. Our approach introduces a trust region in the latent space, restricted to permissible edits (i.e., bounded adjustments along designated semantic directions), and proves that within this region, the sensitivity of attribute scores is bounded. Unlike prior work, we do not assume orthogonality or linear separability of attributes in the latent space. Instead, we bound sensitivity through the Jacobian of the attribute scores evaluated along the chosen subspace, and extend this analysis to joint sub-spaces capturing entangled attributes. Finally, to make privacy certification tractable, we demonstrate that it is sufficient to verify this boundedness on a finite cover of representative points, which guarantees the property over the entire region. On StyleGAN-FFHQ, our empirical evaluation shows that the proposed mechanism lowers protected attribute inference success while preserving visual quality comparable to a DP-calibrated isotropic baseline under the same privacy parameters. Importantly, projection into the protected subspace suppresses non-target leakage: subspace-restricted noise keeps leakage near zero. By concentrating perturbations along protected semantic directions, our method achieves stronger privacy guarantees without sacrificing much visual utility relative to the isotropic noise method.
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.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.004 | 0.012 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.000 | 0.001 |
| Science and technology studies | 0.001 | 0.002 |
| Scholarly communication | 0.002 | 0.004 |
| Open science | 0.002 | 0.003 |
| Research integrity | 0.001 | 0.003 |
| Insufficient payload (model declined to judge) | 0.002 | 0.001 |
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 itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".