pith. sign in

arxiv: 2605.07930 · v1 · submitted 2026-05-08 · 💻 cs.LG · cs.AI

INO-SGD: Addressing Utility Imbalance under Individualized Differential Privacy

Xiao Tian , Jue Fan , Rachael Hwee Ling Sim , Bryan Kian Hsiang Low This is my paper

Pith reviewed 2026-05-11 02:59 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords individualized differential privacyutility imbalancestochastic gradient descentmachine learning privacydata weightingpersonalized privacy
0
0 comments X

The pith

INO-SGD down-weights data with stronger privacy needs inside each SGD batch to reduce utility imbalance while keeping individualized differential privacy intact.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper identifies that individualized differential privacy algorithms create utility imbalance by under-representing data from users with stricter privacy settings, such as sensitive medical cases, which then hurts model accuracy on similar future data. INO-SGD counters this by strategically reducing the influence of those high-privacy points within every training batch so their contribution improves steadily over iterations. The adjustment is constructed so that the per-user privacy guarantees remain satisfied at all times. This matters because data owners now set their own privacy levels, and unbalanced models risk failing precisely on the most protected but important examples.

Core claim

INO-SGD strategically down-weights data within each batch to improve performance on the more private data across all iterations while satisfying IDP. Existing techniques for fixing utility imbalance do not meet IDP constraints and cannot be adapted without losing those guarantees. The method therefore supplies both the imbalance correction and the required privacy property in one algorithm.

What carries the argument

INO-SGD, a stochastic gradient descent variant that assigns per-sample weights inside each batch according to individual privacy levels to counteract under-representation of high-privacy data.

If this is right

  • Trained models achieve higher accuracy on data drawn from the same distribution as the high-privacy subset.
  • The privacy loss for each individual remains bounded exactly as required by their chosen privacy parameter.
  • No separate post-processing or re-weighting stage is needed after the weighted SGD steps.
  • The same batch-wise weighting rule applies uniformly across all training iterations.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The weighting rule could be tested on optimizers other than SGD to check whether the imbalance correction generalizes.
  • In domains with many privacy tiers the method might reduce the need for separate models per privacy level.
  • Deployment pipelines could expose the per-sample weights as an audit log for privacy compliance.

Load-bearing premise

Strategically down-weighting data within each batch will improve performance on more private data across iterations without breaking the individualized privacy guarantees or introducing new imbalances.

What would settle it

A controlled training run in which the down-weighting either produces a measurable privacy violation for some users or leaves accuracy on the high-privacy subset no better than standard IDP-SGD would falsify the central claim.

Figures

Figures reproduced from arXiv: 2605.07930 by Bryan Kian Hsiang Low, Jue Fan, Rachael Hwee Ling Sim, Xiao Tian.

Figure 2
Figure 2. Figure 2: Illustration of IDP￾induced utility imbalance with 2 data owners, BLUE (less private) and RED (more private). Model parame￾ters θ are updated in a direction that reduces BLUE’s loss but increases RED. gradient-based algorithms, which lead to utility imbalance across different groups. In Sec. 3.1 and App. B.3, we explain why this cannot be solved in the same way as existing methods tackling data imbalance. … view at source ↗
Figure 3
Figure 3. Figure 3: Graphical illustration of the INO￾SGD algorithm. At iteration t, a batch Bt (e.g., 10 data) is sampled. The gradients are computed and sorted by descending loss. INO-SGD calculates the average importance score of each gradient by integrating the im￾portance function within its associated interval. By multiplying the clipped gradients to their scores, important gradients are fully kept while less important … view at source ↗
Figure 4
Figure 4. Figure 4: At each iteration t, BIF ft (solid line) is constructed by transforming TIF ftail (blue dashed line). The x-axes refer to the po￾sition of each ordered gradient piece and the y-axes refer to its importance score. When a new datum is added, INO-SGD first ex￾amines if its gradient gd is important based on the rank of d’s loss. If it is deemed important, INO￾SGD simply clips it to its clipping threshold Co(d)… view at source ↗
Figure 5
Figure 5. Figure 5: Per-group utility corresponding to different own￾ers for IDP-SGD and INO-SGD. INO-SGD significantly improves model utility (∼ 10% accuracy) for the more private owners without lowering the utility for the less private owners. 0 2500 5000 7500 9375 No. of iterations t 00 01 02 Validation loss IDP INO (a) MNIST. 0 1000 2000 3000 No. of iterations t 0 20 40 60 Validation acc. IDP INO (b) CIFAR-10. 0 2000 4000… view at source ↗
Figure 6
Figure 6. Figure 6: INO-SGD’s per￾group recall minus IDP-SGD’s. Model utilities for the more pri￾vate owners show higher increase. and BOO as it effectively optimizes a different objective Lftail (see App. C.3.3 for verification that the gradient at each iteration is an unbiased estimate of derivative of Lftail): Theorem 3.5 (Objective of INO-SGD). INO-SGD specified by TIF ftail effectively minimizes Lftail(θ; Db) = 1 K PK k=… view at source ↗
Figure 8
Figure 8. Figure 8: Overview of IDP-induced utility imbalance and the INO-SGD algorithm. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Illustration of DP. When a randomized algorithm A takes in a pair of neighboring datasets, we say A satisfies DP if its output distributions are always close to each other. The close￾ness of distributions is defined differently in different variants of DP. DP works by comparing the distributions of randomized output θ when algorithm A takes in a pair of neighboring datasets D and Dd that differ by only one… view at source ↗
Figure 10
Figure 10. Figure 10: Imbalance of individualized sampling rates and clipping thresholds under IDP￾SGD’s SAMPLE and SCALE variants. (a) and (b) show the discrepancies in sampling rates and clipping thresholds for different privacy preferences. C.1.2 PROOF AND DISCUSSIONS FOR THEOREM 3.1 In this section, we state the formal version of Thm. 3.1, prove it and provide some additional dis￾cussions. Specifically, the first part of t… view at source ↗
Figure 11
Figure 11. Figure 11: Comparison of per-owner model utility between IDP-SGD and INO-SGD. Here O2, 5, 8 refers to Owners 2, 5 and 8 and likewise for the others. INO-SGD consistently improves the model performance for the more private data owners while sometimes conservatively lowering the performance for the less private data owners. In general, it results in a more balanced learning dynamics. 45 [PITH_FULL_IMAGE:figures/full_… view at source ↗
Figure 12
Figure 12. Figure 12: Comparison of worst-owner accuracy between IDP-SGD and INO-SGD. INO￾SGD consistently improves the model performance for the most private data owners and thus ad￾dress utility imbalance. 46 [PITH_FULL_IMAGE:figures/full_fig_p046_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: INO-SGD’s per-group utility minus IDP-SGD’s. Model utilities for the more private owners show higher increase. Another way to assess utility imbalance is through the worst-group accuracy (i.e., the accuracy of the worst group(s) at different stages of training). In [PITH_FULL_IMAGE:figures/full_fig_p047_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Comparison of overall model performance between IDP-SGD and INO-SGD. INO-SGD consistently improves the overall model performance. IDP INO Class 2 Class 4 0 2500 5000 7500 9375 No. of iterations t 0 20 40 60 80 Per-class recall (a) MN-PC O1 C2, 4. IDP INO Class 7 Class 10 0 500 1000 No. of iterations t 0 30 60 90 Per-class recall (b) SU-LS O1 C7, 10. IDP INO Class 0 Class 3 0 2000 4000 No. of iterations t … view at source ↗
Figure 15
Figure 15. Figure 15: Per-class learning dynamics within a data owner. Here O1 C2, 4 refers to Owner 1 Classes 2 and 4, and likewise for the others. Different classes within a data owner are learnt with different dynamics. INO-SGD improves such within-owner learning dynamics too. 48 [PITH_FULL_IMAGE:figures/full_fig_p048_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: INO-SGD is robust to choice of γ for a limited range. After that, it starts to trade overall utility for better balance. In particular, we plot the model performances at 1/4, 2/4, 3/4, and 4/4 of the entire training stage (in terms of iterations). For the CIFAR-10 dataset, we only show 2 data owners for better visualization. (II) α and β in Beta distribution. In [PITH_FULL_IMAGE:figures/full_fig_p049_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: INO-SGD is robust to the choice of α and β if the length of the tail is chosen adequately. The dashed line represents our baseline IDP-SGD. In particular, a small α and large β will cause INO-SGD’s performance to be similar to IDP-SGD and vice versa. (III) Alternative TIF forms. In this section, we consider the alternative function form of tail importance function ftail, a step function, as described in A… view at source ↗
Figure 18
Figure 18. Figure 18: INO-SGD is robust to forms of the tail importance function ftail. In this set of experiments, we use the step function with constant step length as described in App. C.2.5: 1/2 for the first step length, 1/4 for the second step length, 1/8 for the third step length, and 0 for the last step length. Similar trends to our main experiments can be observed. (IV) Impact of order. In [PITH_FULL_IMAGE:figures/fu… view at source ↗
Figure 19
Figure 19. Figure 19: Benefit of using a descending order of loss. D.4.3 PARETO SUPERIORITY Beyond the region where the model owner can simultaneously correct utility imbalance and im￾prove/preserve overall model utility (which only INO-SGD can achieve), the model owner can also choose to set the TIF more aggressively (e.g., downweighting more of the less important gradients by setting a larger tail length γ) in order to trade… view at source ↗
Figure 20
Figure 20. Figure 20: Pareto superiority of INO-SGD. ▲ denotes IDP-SGD and ▼ denotes DP-SGD using the strongest privacy. ⋆ shows our reported results where the model owner attains the largest dual improvement in both utility balance and overall utility. The red arrow indicates that INO-SGD Pareto-dominates any method whose tradeoff curve lies to the upper left of INO-SGD’s, including the simple baseline. Therefore, INO-SGD off… view at source ↗
Figure 21
Figure 21. Figure 21: INO-SGD is highly compatible with adaptive clipping. (d) shows that for adaptive clipping Owner 3 suffers from MID till the end, which verifies our theoretical analysis that adaptive clipping could lengthen IDP-induced MID. Since INO-SGD addresses MID and utility imbalance, it significantly improves the performance of such methods. D.4.5 PRIVACY ASSESSMENT VIA MEMBERSHIP INFERENCE ATTACK In Sec. 3.3.1 and… view at source ↗
Figure 22
Figure 22. Figure 22: The privacy of INO-SGD is validated by LiRA membership inference attack. The AUROC for all owners are close to 0.5 and the ROC curves for IDP-SGD and INO-SGD are similar. The AUROCs for the more private owner (smaller ϵ) and less private owner (larger ϵ) are respectively (a) 0.571, 0.645, (b) 0.563, 0.663, (c) 0.569, 0.623, and (d) 0.563, 0.580 for the subfigures. E ADDITIONAL DISCUSSIONS E.1 LIMITATIONS … view at source ↗
Figure 23
Figure 23. Figure 23: Intuitive illustration of the IDP-balance-utility tradeoff. The privacy budgets of two owners constrain how much the model may use their data, so the trained model will either under￾utilize data from the less private owners and under-perform (left), or utilize data from both owners unevenly which causes utility imbalance (right). F OTHER QUESTIONS 1. Why do we consider the individualized privacy setting? … view at source ↗
read the original abstract

Differential privacy (DP) is widely employed in machine learning to protect confidential or sensitive training data from being revealed. As data owners gain greater control over their data due to personal data ownership, they are more likely to set their own privacy requirements, necessitating individualized DP (IDP) to fulfil such requests. In particular, owners of data from more sensitive subsets, such as positive cases of stigmatized diseases, likely set stronger privacy requirements, as leakage of such data could incur more serious societal impact. However, existing IDP algorithms induce a critical utility imbalance problem: Data from owners with stronger privacy requirements may be severely underrepresented in the trained model, resulting in poorer performance on similar data from subsequent users during deployment. In this paper, we analyze this problem and propose the INO-SGD algorithm, which strategically down-weights data within each batch to improve performance on the more private data across all iterations. Notably, our algorithm is specially designed to satisfy IDP, while existing techniques addressing utility imbalance neither satisfy IDP nor can be easily adapted to do so. Lastly, we demonstrate the empirical feasibility of our approach.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes INO-SGD to address utility imbalance under individualized differential privacy (IDP). It observes that data owners with stronger privacy requirements (e.g., sensitive medical cases) cause their samples to be underrepresented in the trained model under standard IDP mechanisms. The algorithm strategically down-weights samples within each batch to boost performance on high-privacy data across iterations, while claiming to be specially designed to satisfy IDP—unlike prior imbalance-correction techniques that neither meet IDP nor adapt easily to it. Empirical results are presented to demonstrate feasibility.

Significance. If the IDP invariant is rigorously established and the utility gains are shown without new privacy violations, the work would meaningfully advance practical deployment of DP in heterogeneous-privacy settings such as healthcare or personalized services. The explicit focus on IDP compliance by design, rather than post-hoc fixes, distinguishes it from existing literature on utility imbalance.

major comments (3)
  1. [Algorithm and Privacy Analysis sections] The core claim that INO-SGD satisfies IDP while performing batch down-weighting requires that the per-sample sensitivity and noise calibration explicitly incorporate the weights w_i (e.g., noise scaled to max |w_i · clipped ∇ℓ_i| rather than the unweighted clipping bound). No equations or accountant adjustment are visible in the abstract or high-level description to confirm this folding; without it, the privacy loss for low-ε samples can exceed their budget when their relative weight increases.
  2. [Introduction and Related Work] The assertion that existing imbalance techniques 'neither satisfy IDP nor can be easily adapted' is load-bearing for the novelty claim. The manuscript must demonstrate (via a concrete counter-example or failed adaptation attempt) why standard re-weighting or re-sampling methods cannot be made IDP-compliant by simply adjusting the noise multiplier, rather than asserting non-adaptability at a high level.
  3. [INO-SGD Algorithm description] The down-weighting rule is described as 'strategically' chosen to improve performance on more private data, yet the selection of w_i appears deterministic from the known privacy vector. It is unclear whether this choice preserves the 'individualized' property across iterations or introduces a new form of imbalance; a formal statement of the weight function and its effect on the gradient expectation is needed.
minor comments (2)
  1. [Abstract] The abstract states that the approach 'demonstrate[s] the empirical feasibility' but provides no details on datasets, baselines, or metrics; a short summary of the experimental setup would improve readability.
  2. [Preliminaries and Algorithm] Notation for the privacy vector, per-sample weights, and the resulting sensitivity bound should be introduced consistently in the algorithm section to aid readers unfamiliar with IDP.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the thoughtful and constructive report. The comments highlight important points on clarity of the privacy analysis, the novelty argument, and the formalization of the weighting scheme. We address each major comment below and will incorporate revisions to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Algorithm and Privacy Analysis sections] The core claim that INO-SGD satisfies IDP while performing batch down-weighting requires that the per-sample sensitivity and noise calibration explicitly incorporate the weights w_i (e.g., noise scaled to max |w_i · clipped ∇ℓ_i| rather than the unweighted clipping bound). No equations or accountant adjustment are visible in the abstract or high-level description to confirm this folding; without it, the privacy loss for low-ε samples can exceed their budget when their relative weight increases.

    Authors: We agree that explicit incorporation of the weights into the sensitivity bound and noise calibration is essential for the IDP guarantee. The full manuscript (Section 4 and Theorem 1) defines the per-sample noise scale as σ_i = (C · w_i) / ε_i, where the weighted gradient norm is used in the sensitivity calculation and the moments accountant is applied per sample. To improve visibility, we will add the explicit weighted sensitivity equation and a short accountant adjustment paragraph to the high-level algorithm description and introduction. revision: yes

  2. Referee: [Introduction and Related Work] The assertion that existing imbalance techniques 'neither satisfy IDP nor can be easily adapted' is load-bearing for the novelty claim. The manuscript must demonstrate (via a concrete counter-example or failed adaptation attempt) why standard re-weighting or re-sampling methods cannot be made IDP-compliant by simply adjusting the noise multiplier, rather than asserting non-adaptability at a high level.

    Authors: The referee correctly identifies that a concrete demonstration would strengthen the novelty claim. We will add a brief counter-example in the Related Work section (or an appendix) showing that applying standard re-weighting to IDP-SGD without per-sample noise recalibration causes the effective privacy loss for low-ε samples to exceed their budget, computed via the weighted moments accountant. This illustrates why simple multiplier adjustment is insufficient. revision: yes

  3. Referee: [INO-SGD Algorithm description] The down-weighting rule is described as 'strategically' chosen to improve performance on more private data, yet the selection of w_i appears deterministic from the known privacy vector. It is unclear whether this choice preserves the 'individualized' property across iterations or introduces a new form of imbalance; a formal statement of the weight function and its effect on the gradient expectation is needed.

    Authors: We will insert a formal definition of the weight function: w_i = ε_i / max_{j in batch} ε_j (normalized to preserve batch sum). Because each w_i is computed from the fixed, known per-sample privacy vector and the noise is calibrated individually, the IDP guarantee is preserved across iterations. We will also add a short lemma proving that the weighted gradient estimator remains unbiased in expectation after batch normalization. revision: yes

Circularity Check

0 steps flagged

No circularity detected; derivation chain self-contained with no reductions to inputs by construction

full rationale

The provided abstract and claims introduce INO-SGD as a new algorithm that down-weights batches to address utility imbalance while satisfying IDP, with the explicit statement that existing imbalance techniques neither satisfy IDP nor adapt easily. No equations, parameter fits, predictions, or derivations are shown that reduce to the inputs by construction. No self-citations, uniqueness theorems, or ansatzes are invoked in the text. The central claim is a design proposal whose validity rests on external verification of the IDP accounting rather than any internal tautology or renaming of known results. This is the normal case of a self-contained algorithmic contribution.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Based solely on abstract; no explicit free parameters, axioms, or invented entities are described in the provided text.

pith-pipeline@v0.9.0 · 5502 in / 1023 out tokens · 23660 ms · 2026-05-11T02:59:21.012364+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

111 extracted references · 111 canonical work pages

  1. [1]

    Journal of Privacy and Confidentiality , volume=

    Heterogeneous Differential Privacy , author=. Journal of Privacy and Confidentiality , volume=

    work page

  2. [2]

    IEEE Transactions on Neural Networks , volume=

    An Improved Algorithm for Neural Network Classification of Imbalanced Training Sets , author=. IEEE Transactions on Neural Networks , volume=. 1993 , publisher=

    work page 1993

  3. [3]

    Journal of Privacy and Confidentiality , volume=

    Privacy Profiles and Amplification by Subsampling , author=. Journal of Privacy and Confidentiality , volume=

    work page

  4. [4]

    American Sociological Review , volume=

    The Stigma of Diseases: Unequal Burden, Uneven Decline , author=. American Sociological Review , volume=. 2023 , publisher=

    work page 2023

  5. [5]

    ACM Computing Surveys , volume=

    Fairness in Machine Learning: A Survey , author=. ACM Computing Surveys , volume=. 2024 , publisher=

    work page 2024

  6. [6]

    The economic journal , volume=

    The Measurement of the Inequality of Incomes , author=. The economic journal , volume=. 1920 , publisher=

    work page 1920

  7. [7]

    cultural fairness

    Another look at “cultural fairness” , author=. Journal of educational measurement , volume=. 1971 , publisher=

    work page 1971

  8. [8]

    Operations research , volume=

    Distributionally Robust Optimization and Its Tractable Approximations , author=. Operations research , volume=. 2010 , publisher=

    work page 2010

  9. [9]

    IEEE Transactions on Knowledge and Data Engineering , volume=

    Learning from Imbalanced Data , author=. IEEE Transactions on Knowledge and Data Engineering , volume=. 2009 , publisher=

    work page 2009

  10. [10]

    2022 , journal=

    Examining the Intersection of Data Privacy and Civil Rights , author=. 2022 , journal=

    work page 2022

  11. [11]

    Differentially private synthetic data via foundation model

    Lin, Zinan and Gopi, Sivakanth and Kulkarni, Janardhan and Nori, Harsha and Yekhanin, Sergey , journal=. Differentially private synthetic data via foundation model

    work page

  12. [12]

    ACM Transactions on Knowledge Discovery From Data , volume=

    l -Diversity: Privacy Beyond k -Anonymity , author=. ACM Transactions on Knowledge Discovery From Data , volume=. 2007 , publisher=

    work page 2007

  13. [13]

    Cell , volume=

    Identifying medical diagnoses and treatable diseases by image-based deep learning , author=. Cell , volume=. 2018 , publisher=

    work page 2018

  14. [14]

    2009 , journal=

    Learning Multiple Layers of Features from Tiny Images , author=. 2009 , journal=

    work page 2009

  15. [15]

    Proceedings of the IEEE , volume=

    Gradient-Based Learning Applied to Document Recognition , author=. Proceedings of the IEEE , volume=. 1998 , publisher=

    work page 1998

  16. [16]

    Incentivizing the Sharing of Healthcare Data in the

    Andreas Panagopoulos and Timo Minssen and Katerina Sideri and Helen Yu and Marcelo Corrales Compagnucci , keywords =. Incentivizing the Sharing of Healthcare Data in the. Computer Law & Security Review , volume =. 2022 , issn =. doi:https://doi.org/10.1016/j.clsr.2022.105670 , url =

    work page doi:10.1016/j.clsr.2022.105670 2022

  17. [17]

    Pardau, Stuart L , journal=. The. 2018 , publisher=

    work page 2018

  18. [18]

    ACM Computing Surveys (CSUR) , volume=

    A Review on Fairness in Machine Learning , author=. ACM Computing Surveys (CSUR) , volume=. 2022 , publisher=

    work page 2022

  19. [19]

    1912 , publisher=

    Wealth and welfare , author=. 1912 , publisher=

    work page 1912

  20. [20]

    Regulation, General Data Protection , journal=. General

    work page

  21. [21]

    Robertson, Sean , journal=

    work page

  22. [22]

    Carnegie Mellon University, Data Privacy , year=

    Simple Demographics Often Identify People Uniquely , author=. Carnegie Mellon University, Data Privacy , year=

    work page

  23. [23]

    International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems , volume=

    k -Anonymity: A Model for Protecting Privacy , author=. International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems , volume=. 2002 , publisher=

    work page 2002

  24. [24]

    2014 , publisher=

    Van Erven, Tim and Harremos, Peter , journal=. 2014 , publisher=

    work page 2014

  25. [25]

    Generalized

    Weymark, John A , journal=. Generalized. 1981 , publisher=

    work page 1981

  26. [26]

    IEEE Transactions on Neural Networks and Learning Systems , volume=

    Balancing Learning Model Privacy, Fairness, and Accuracy with Early Stopping Criteria , author=. IEEE Transactions on Neural Networks and Learning Systems , volume=. 2021 , publisher=

    work page 2021

  27. [27]

    International Journal of Human-Computer Studies , volume=

    Privacy practices of Internet users: Self-reports versus observed behavior , author=. International Journal of Human-Computer Studies , volume=. 2005 , publisher=

    work page 2005

  28. [28]

    actual behavior , author=

    Privacy in e-commerce: Stated preferences vs. actual behavior , author=. Communications of the ACM , volume=. 2005 , publisher=

    work page 2005

  29. [29]

    Journal of Machine Learning Research , volume=

    Inherent tradeoffs in learning fair representations , author=. Journal of Machine Learning Research , volume=

    work page

  30. [30]

    Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security , pages=

    Deep Learning with Differential Privacy , author=. Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security , pages=

    work page 2016

  31. [31]

    On k -Anonymity and the Curse of Dimensionality , author=. Proc. VLDB , volume=

    work page

  32. [32]

    Faster rates of convergence to stationary points in differentially private optimization , author=. Proc. ICML , pages=. 2023 , organization=

    work page 2023

  33. [33]

    Differential Privacy Has Disparate Impact on Model Accuracy , author=. Proc. NeurIPS , volume=

    work page

  34. [34]

    Hypothesis Testing Interpretations and

    Balle, Borja and Barthe, Gilles and Gaboardi, Marco and Hsu, Justin and Sato, Tetsuya , booktitle=. Hypothesis Testing Interpretations and

    work page

  35. [35]

    Private stochastic convex optimization with optimal rates , author=. Proc. NeurIPS , volume=

    work page

  36. [36]

    Have It Your Way: Individualized Privacy Assignment for

    Boenisch, Franziska and M. Have It Your Way: Individualized Privacy Assignment for. Proc. NeurIPS , volume=

    work page

  37. [37]

    IEEE Symposium on Security and Privacy , pages=

    Machine Unlearning , author=. IEEE Symposium on Security and Privacy , pages=

    work page

  38. [38]

    Scalable and efficient training of large convolutional neural networks with differential privacy , author=. Proc. NeurIPS , volume=

    work page

  39. [39]

    Automatic clipping: Differentially private deep learning made easier and stronger , author=. Proc. NeurIPS , volume=

    work page

  40. [40]

    2022 IEEE symposium on security and privacy (SP) , pages=

    Membership inference attacks from first principles , author=. 2022 IEEE symposium on security and privacy (SP) , pages=. 2022 , organization=

    work page 2022

  41. [41]

    Big self-supervised models are strong semi-supervised learners , author=. Proc. NeurIPS , volume=

    work page

  42. [42]

    Understanding gradient clipping in private

    Chen, Xiangyi and Wu, Steven Z and Hong, Mingyi , booktitle=. Understanding gradient clipping in private

    work page

  43. [43]

    Chen, Yongqiang and Zhou, Kaiwen and Bian, Yatao and Xie, Binghui and Wu, Bingzhe and Zhang, Yonggang and KAILI, MA and Yang, Han and Zhao, Peilin and Han, Bo and others , booktitle=

    work page

  44. [44]

    Beyond uniform

    Das, Rudrajit and Kale, Satyen and Xu, Zheng and Zhang, Tong and Sanghavi, Sujay , booktitle=. Beyond uniform. 2023 , organization=

    work page 2023

  45. [45]

    Differentially private and fair classification via calibrated functional mechanism , author=. Proc. AAAI , volume=

    work page

  46. [46]

    Retiring

    Ding, Frances and Hardt, Moritz and Miller, John and Schmidt, Ludwig , booktitle=. Retiring

    work page

  47. [47]

    Dukler, Yonatan and Bowman, Benjamin and Achille, Alessandro and Golatkar, Aditya and Swaminathan, Ashwin and Soatto, Stefano , booktitle=

    work page

  48. [48]

    Is Fairness Only Metric Deep? Evaluating and Addressing Subgroup Gaps in Deep Metric Learning , author=. Proc. ICLR , year=

    work page

  49. [49]

    International Colloquium on Automata, Languages, and Programming , pages=

    Differential Privacy , author=. International Colloquium on Automata, Languages, and Programming , pages=

    work page

  50. [50]

    Advances in Cryptology-EUROCRYPT 2006: 24th Annual International Conference on the Theory and Applications of Cryptographic Techniques , pages=

    Our Data, Ourselves: Privacy via Distributed Noise Generation , author=. Advances in Cryptology-EUROCRYPT 2006: 24th Annual International Conference on the Theory and Applications of Cryptographic Techniques , pages=. 2006 , organization=

    work page 2006

  51. [51]

    IEEE 51st Annual Symposium on Foundations of Computer Science , pages=

    Boosting and Differential Privacy , author=. IEEE 51st Annual Symposium on Foundations of Computer Science , pages=. 2010 , organization=

    work page 2010

  52. [52]

    Proceedings of the 3rd Innovations in Theoretical Computer Science Conference , pages=

    Fairness through awareness , author=. Proceedings of the 3rd Innovations in Theoretical Computer Science Conference , pages=

    work page

  53. [53]

    Disparate Impact in Differential Privacy from Gradient Misalignment , author=. Proc. ICLR , year=

    work page

  54. [54]

    Improved Convergence of Differential Private

    Fang, Huang and Li, Xiaoyun and Fan, Chenglin and Li, Ping , booktitle=. Improved Convergence of Differential Private

    work page

  55. [55]

    Proceedings of the 2020 Workshop on Privacy-Preserving Machine Learning in Practice , pages=

    Neither private nor fair: Impact of data imbalance on utility and fairness in differential privacy , author=. Proceedings of the 2020 Workshop on Privacy-Preserving Machine Learning in Practice , pages=

    work page 2020

  56. [56]

    A Theoretical Analysis of the Learning Dynamics under Class Imbalance , author=. Proc. ICML , pages=. 2023 , organization=

    work page 2023

  57. [57]

    Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency , pages=

    On the Impact of Machine Larning Randomness on Group Fairness , author=. Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency , pages=

    work page 2023

  58. [58]

    2022 , organization=

    Ganev, Georgi and Oprisanu, Bristena and De Cristofaro, Emiliano , booktitle=. 2022 , organization=

    work page 2022

  59. [59]

    Equality of opportunity in supervised learning , author=. Proc. NeurIPS , volume=

    work page

  60. [60]

    Deep Residual Learning for Image Recognition , author=. Proc. CVPR , pages=

    work page

  61. [61]

    Differentially private fair learning , author=. Proc. ICML , pages=. 2019 , organization=

    work page 2019

  62. [62]

    Proceedings of the International Conference on Artificial Intelligence , volume=

    The Class Imbalance Problem: Significance and Strategies , author=. Proceedings of the International Conference on Artificial Intelligence , volume=

    work page

  63. [63]

    Kawaguchi, Kenji and Lu, Haihao , booktitle=. Ordered. 2020 , organization=

    work page 2020

  64. [64]

    Imagenet Classification with Deep Convolutional Neural Networks , author=. Proc. NeurIPS , volume=

    work page

  65. [65]

    What you see is what you get: Principled deep learning via distributional generalization , author=. Proc. NeurIPS , volume=

    work page

  66. [66]

    Proceedings of the IEEE international conference on computer vision , pages=

    Deeper, Broader and Artier Domain Generalization , author=. Proceedings of the IEEE international conference on computer vision , pages=

    work page

  67. [67]

    Convergence of

    Li, Haochuan and Rakhlin, Alexander and Jadbabaie, Ali , booktitle=. Convergence of

    work page

  68. [68]

    Workshop on Algorithmic Fairness through the Lens of Causality and Privacy , pages=

    Stochastic differentially private and fair learning , author=. Workshop on Algorithmic Fairness through the Lens of Causality and Privacy , pages=. 2023 , organization=

    work page 2023

  69. [69]

    51st Annual Allerton Conference on Communication, Control, and Computing , pages=

    Privacy-Utility Tradeoff Under Statistical Uncertainty , author=. 51st Annual Allerton Conference on Communication, Control, and Computing , pages=. 2013 , organization=

    work page 2013

  70. [70]

    Mitigating Disparate Impact of Differential Privacy in Federated Learning through Robust Clustering , author=. Proc. CVPR Causal and Object-Centric Representations for Robotics Workshop , year=

    work page

  71. [71]

    Mironov, Ilya , booktitle=

    work page

  72. [72]

    Nearly tight black-box auditing of differentially private machine learning , author=. Proc. NeurIPS , volume=

    work page

  73. [73]

    Proceedings of the 2nd International Conference on Computing Advancements , pages=

    Class Imbalance Problems in Machine Learning: A Review of Methods and Future Challenges , author=. Proceedings of the 2nd International Conference on Computing Advancements , pages=

    work page

  74. [74]

    Annals of Operations Research , volume=

    Multiple Criteria Linear Programming Model for Portfolio Selection , author=. Annals of Operations Research , volume=. 2000 , publisher=

    work page 2000

  75. [75]

    Tempered Sigmoid Activations for Deep Learning with Differential Privacy , author=. Proc. AAAI , volume=

    work page

  76. [76]

    Group-robust Sample Reweighting for Subpopulation Shifts via Influence Functions , author=. Proc. ICLR , year=

    work page

  77. [77]

    Focal loss for dense object detection , author=. Proc. ICCV , pages=

    work page

  78. [78]

    Minimizing the Maximal Loss: How and why , author=. Proc. ICML , pages=. 2016 , organization=

    work page 2016

  79. [79]

    IEEE Symposium on Security and Privacy , pages=

    Membership Inference Attacks Against Machine Learning Models , author=. IEEE Symposium on Security and Privacy , pages=. 2017 , organization=

    work page 2017

  80. [80]

    Training Region-Based Object Detectors with Online Hard Example Mining , author=. Proc. CVPR , pages=

    work page

Showing first 80 references.