mlsanityData to 7 Oct 2026

Adversarial Sample-Based Approach for Tighter Privacy Auditing in Final Model-Only Scenarios

arXiv 2412.01756 ↗First posted 2 Dec 2024cs.CR cs.LGcited by 0

Not on the map. Why?

Abstract (arXiv)

Auditing Differentially Private Stochastic Gradient Descent (DP-SGD) in the final model setting is challenging and often results in empirical lower bounds that are significantly looser than theoretical privacy guarantees. We introduce a novel auditing method that achieves tighter empirical lower bounds without additional assumptions by crafting worst-case adversarial samples through loss-based input-space auditing. Our approach surpasses traditional canary-based heuristics and is effective in final model-only scenarios. Specifically, with a theoretical privacy budget of varepsilon = 10.0, our method achieves empirical lower bounds of 4.914, compared to the baseline of 4.385 for MNIST. Our work offers a practical framework for reliable and accurate privacy auditing in differentially private machine learning.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Builds on 36 of its 48 references that are in the map. First posted on arXiv 2 Dec 2024.

Builds on

36 of 48 references.

References · 36

  1. 2016Deep Learning with Differential Privacycited by 2,682
  2. 2016Membership Inference Attacks against Machine Learning Modelscited by 2,221
  3. 2018The Secret Sharer: Evaluating and Testing Unintended Memorization in Neural Networkscited by 847
  4. 2021Membership Inference Attacks From First Principlescited by 839
  5. 2017Renyi Differential Privacycited by 643
  6. 2021Opacus: User-Friendly Differential Privacy Library in PyTorchcited by 375
  7. 2016Semi-supervised Knowledge Transfer for Deep Learning from Private Training Datacited by 347
  8. 2013The Composition Theorem for Differential Privacycited by 290
  9. 2022Unlocking High-Accuracy Differentially Private Image Classification through Scalecited by 218
  10. 2020Auditing Differentially Private Machine Learning: How Private is Private SGD?cited by 213
  11. 2021Adversary Instantiation: Lower Bounds for Differentially Private Machine Learningcited by 193
  12. 2021Numerical Composition of Differential Privacycited by 174
24 more
  1. 2022Reconstructing Training Data with Informed Adversariescited by 143
  2. 2022Reconstructing Training Data from Trained Neural Networkscited by 129
  3. 2023Privacy Auditing with One (1) Training Runcited by 119
  4. 2020Synthetic Data -- Anonymisation Groundhog Daycited by 105
  5. 2018Privacy Amplification by Iterationcited by 97
  6. 2021Differential Privacy Dynamics of Langevin Diffusion and Noisy Gradient Descentcited by 71
  7. 2022Privacy of Noisy Stochastic Gradient Descent: More Iterations without More Privacy Losscited by 60
  8. 2024Position: LLM Unlearning Benchmarks are Weak Measures of Progresscited by 58
  9. 2022Connect the Dots: Tighter Discrete Approximations of Privacy Loss Distributionscited by 50
  10. 2022Differentially Private Learning Needs Hidden State (Or Much Faster Convergence)cited by 45
  11. 2022Debugging Differential Privacy: A Case Study for Privacy Auditingcited by 39
  12. 2018Detecting Violations of Differential Privacycited by 37
  13. 2023One-shot Empirical Privacy Estimation for Federated Learningcited by 35
  14. 2024Evaluations of Machine Learning Privacy Defenses are Misleadingcited by 35
  15. 2022Bayesian Estimation of Differential Privacycited by 31
  16. 2024Auditing f-Differential Privacy in One Runcited by 28
  17. 2022CANIFE: Crafting Canaries for Empirical Privacy Measurement in Federated Learningcited by 24
  18. 2019Privacy Amplification by Mixing and Diffusion Mechanismscited by 21
  19. 2024Nearly Tight Black-Box Auditing of Differentially Private Machine Learningcited by 19
  20. 2024Tighter Privacy Auditing of DP-SGD in the Hidden State Threat Modelcited by 17
  21. 2024Shifted Interpolation for Differential Privacycited by 10
  22. 2024It's Our Loss: No Privacy Amplification for Hidden State DP-SGD With Non-Convex Losscited by 10
  23. 2024The Last Iterate Advantage: Empirical Auditing and Principled Heuristic Analysis of Differentially Private SGDcited by 10
  24. 2024Auditing Private Predictioncited by 7

Built on by

Not cited yet.