The DOI matches its Crossref record; 80 of the 80 references it cites were found in a registry or an index; 14 of the 28 sources its claims cite were checked for support.
When Machine Unlearning Jeopardizes Privacy
Open accessAbstract
This paper performs the first study on unintended information leakage caused by machine unlearning and proposes a novel membership inference attack.
Overview
CoChat Summary
## TL;DR
The paper shows that machine unlearning can unintentionally leak membership information by letting an attacker compare outputs of the original and unlearned models. Across multiple datasets and model types, the proposed attack often outperforms classical membership inference, especially on well-generalized models.
## Key findings
- On categorical datasets, the attack “consistently better[s] than classical membership inference” and achieves up to 0.48 improvement in AUC; the best reported DegCount and DegRate are 0.94 and 0.40.
- On image datasets, the attack also outperforms classical membership inference in all settings; for CIFAR10 + DenseNet, the paper reports an overfitting level of 0.439, with the highest privacy degradation, while MNIST + SimpleCNN has overfitting < 0.05 and the lowest degradation.
- For a well-generalized decision tree on the Adult dataset, classical membership inference is near random with AUC = 0.497, while the proposed attack reaches AUC = 0.882; the paper also reports DegCount = 0.85 and DegRate = 0.28 for this example.
- Transfer is strong: in one dataset-transfer example, the attack AUC drops only 1% (from 0.944 to 0.931) when moving from Insta-NY → Insta-LA for a decision tree target model.
- Against the approximate SISA unlearning method, attack performance drops compared with retraining from scratch, indicating SISA reduces the influence of a specific sample on the global model.
## Methods
- The paper studies black-box membership inference in machine unlearning: the adversary queries both the original model and the unlearned model, extracts their posterior outputs, constructs features from both, and trains a binary attack model.
- It evaluates on categorical datasets (Adult, Accident, Insta-NY, Insta-LA) and image datasets (MNIST, CIFAR10, STL10) using target models including logistic regression, decision tree, random forest, MLP, SimpleCNN, DenseNet, and ResNet50.
- Main analysis compares feature constructions (DirectConcat, SortedConcat, DirectDiff, SortedDiff, EucDist), tests overfitting effects, shadow-model transferability, multiple intermediate unlearned models, group deletion, online learning, and defense mechanisms.
## Limitations
- The paper explicitly notes that the evaluation of the attack in other scenarios is deferred to Section 6, and that some experiments are limited by space; for example, the multiple-intermediate-models experiment is concentrated on Insta-NY with three target models.
- The defense DP-SGD is only reported for logistic regression because it can only be applied to models that undergo gradient updating; this is stated in the paper.
- Evident from the text: several reported results are based on selected datasets/models and some figures/tables are partially summarized in the excerpt, so the paper’s broader empirical coverage is not fully visible here.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
5 Valid · 1 Not valid · 5 Not checked · 94 Own results and interpretations
Every source the paper cites for this claim was checked and supports it.
In the paper
“This is consistent with the conclusion of previous studies [60, 64] that classical membership inference (which uses plain posterior information) performs well on overfitted target models.”
What the sources say
- Supportedcites [60] [60] Ahmed Salem, Yang Zhang, Mathias Humbert, Pascal Berrang, Mario Fritz, and Michael Backes. ML-Leaks: Model and Data Independent Membership Inference Attacks and Defenses on Machine Learning Models. In Network and Distributed System Security Symposium (NDSS) . Internet Society, 2019. 1, 3, 4, 6, 8, 13, 14 - source
The Experimental Setup (page not provided) reports that more-overfitted target models are more vulnerable to membership inference, with substantially stronger attack performance on a highly overfitted model. The Membership Inference section (page not provided) describes an attack that uses the target model’s output posteriors to predict membership. (LLM, conf=0.91)
- Supportedcites [64] [64] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership Inference Attacks Against Machine Learning Models. In IEEE Symposium on Security and Privacy (S&P) , pages 3-18. IEEE, 2017. 1, 3, 4, 8, 12, 13, 14 - source
The attack overview (p. 4) says the membership-inference attack uses the target model’s prediction vector of class probabilities. “Effect of overfitting” (p. 11) reports that larger train–test accuracy gaps are associated with higher membership-inference precision. The paper also cautions that overfitting is not the only factor affecting attack success. (LLM, conf=0.92)
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“For the well-generalized target models, exploiting the plain posterior information has shown to perform poorly in terms of membership inference [60, 64].”
What the sources say
- Supportedcites [60] [60] Ahmed Salem, Yang Zhang, Mathias Humbert, Pascal Berrang, Mario Fritz, and Michael Backes. ML-Leaks: Model and Data Independent Membership Inference Attacks and Defenses on Machine Learning Models. In Network and Distributed System Security Symposium (NDSS) . Internet Society, 2019. 1, 3, 4, 6, 8, 13, 14 - source
The Experimental Setup discussion of Figure 2 reports that models with less overfitting are less vulnerable to membership inference: the Adult target model had only a 2% training–testing accuracy gap, and the attack achieved about 0.5 precision and recall. The paper describes this attack as using the target model’s output posteriors. (LLM, conf=0.87)
- Partly supportedcites [64] [64] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership Inference Attacks Against Machine Learning Models. In IEEE Symposium on Security and Privacy (S&P) , pages 3-18. IEEE, 2017. 1, 3, 4, 8, 12, 13, 14 - sourceExperimental — accuracy not yet measured
Section F (p. 11) reports near-random membership-inference precision for a target model with nearly identical training and test accuracy, and Section VII (p. 12) links attack success to generalizability. But the paper evaluates a trained attack model using prediction outputs, not a separate method that exploits only plain posterior information; it also says overfitting is not the only factor affecting leakage. (LLM, conf=0.84)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“More importantly, we show that our attack in multiple cases outperforms the classical membership inference attack on the original ML model, which indicates that machine unlearning can have counterproductive effects on privacy.”
boostedstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Rests on 5 assumptions not shown here.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“We notice that the privacy degradation is especially significant for well-generalized ML models where classical membership inference does not perform well.”
- Also stated in: § When Machine Unlearning Jeopardizes Privacy · plain
“In particular, we observe that privacy is especially degraded because of machine unlearning in the case of well-generalized models.”
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Rests on 5 assumptions not shown here.
- Also stated in: § When Machine Unlearning Jeopardizes Privacy · plain
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“We further investigate four mechanisms to mitigate the newly discovered privacy risks and show that releasing the predicted label only, temperature scaling, and differential privacy are effective.”
- Also stated in: § When Machine Unlearning Jeopardizes Privacy · plain
“publishing only the predicted label, temperature scaling, and differential privacy can effectively prevent our attack.”
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Rests on 5 assumptions not shown here.
- Also stated in: § When Machine Unlearning Jeopardizes Privacy · plain
58 more claims
Create a free account to see all 105 claims and to verify the 7 references still open.
- Probably not a substantial claim. 16
- It may state two separate claims as one. 28
- The quote only partly states this claim. 2
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Thus, a membership inference attack or a poisoning attack would reveal a significant amount of detailed information about the unlearned samples [78, 108].”
From Machine Unlearning: A Survey (2024) · cites this paper as [78]
Experimental — accuracy not yet measuredThe cited source says what the claim attributes to it.
In the citing paper
“This could exacerbate user privacy issues if an attacker has access to the model before and after the unlearning operation [78].”
From Machine Unlearning: A Survey (2024) · cites this paper as [78]
Citations
Cites
Deep Residual Learning for Image Recognition
He, K., Zhang, X., Ren, S., & Sun, J. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.90 · Cited by 235,474
Densely Connected Convolutional Networks
Huang, G., Liu, Z., Van Der Maaten, L., & Weinberger, K. Q. (2017) · 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2017.243 · Cited by 48,566
Deep Learning with Differential Privacy
Abadi, M., Chu, A., Goodfellow, I., McMahan, H. B., Mironov, I., Talwar, K., & Zhang, L. (2016) · Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security · doi:10.1145/2976749.2978318 · Cited by 6,578
The Algorithmic Foundations of Differential Privacy
Dwork, C., & Roth, A. (2014) · Foundations and Trends® in Theoretical Computer Science · doi:10.1561/0400000042 · Cited by 5,557
Membership Inference Attacks Against Machine Learning Models
Shokri, R., Stronati, M., Song, C., & Shmatikov, V. (2017) · 2017 IEEE Symposium on Security and Privacy (SP) · doi:10.1109/sp.2017.41 · Cited by 4,754 · cited 10×
“, Po and Pu , also referred to as confidence values [64].”
BackgroundMethodResultInfluentialThe Limitations of Deep Learning in Adversarial Settings
Papernot, N., McDaniel, P., Jha, S., Fredrikson, M., Celik, Z. B., & Swami, A. (2016) · 2016 IEEE European Symposium on Security and Privacy (EuroS&P) · doi:10.1109/eurosp.2016.36 · Cited by 3,993
The Algorithmic Foundations of Differential Privacy
Dwork, C., & Roth, A. (2013) · doi:10.1561/9781601988195 · Cited by 3,856
Practical Black-Box Attacks against Machine Learning
Papernot, N., McDaniel, P., Goodfellow, I., Jha, S., Celik, Z. B., & Swami, A. (2017) · Proceedings of the 2017 ACM on Asia Conference on Computer and Communications Security · doi:10.1145/3052973.3053009 · Cited by 3,495
On calibration of modern neural networks
Guo, C., Pleiss, G., Sun, Y., & Weinberger, K. Q. (2017) · arXiv · doi:10.48550/arxiv.1706.04599 · Cited by 3,435
Model Inversion Attacks that Exploit Confidence Information and Basic Countermeasures
Fredrikson, M., Jha, S., & Ristenpart, T. (2015) · Proceedings of the 22nd ACM SIGSAC Conference on Computer and Communications Security · doi:10.1145/2810103.2813677 · Cited by 2,961
Research works printed but not linked: 1
- Reza Shokri, Martin Strobel, and Yair Zick. Exploiting Transparency Measures for Membership Inference: a Cautionary Tale. In The AAAI Workshop on Privacy-Preserving Artificial Intelligence (PPAI). AAAI, 2020.
Cited by
Fast Yet Effective Machine Unlearning
Tarun, A. K., Chundawat, V. S., Mandal, M., & Kankanhalli, M. (2021) · arXiv · doi:10.48550/arxiv.2111.08947 · Cited by 190
Fast Yet Effective Machine Unlearning
Tarun, A. K., Chundawat, V. S., Mandal, M., & Kankanhalli, M. (2024) · IEEE Transactions on Neural Networks and Learning Systems · doi:10.1109/tnnls.2023.3266233 · Cited by 182
“Some researchers have studied the unintended privacy risks resulting from the existing unlearning methods [34], [35].”
BackgroundChundawat, V. S., Tarun, A. K., Mandal, M., & Kankanhalli, M. (2023) · IEEE Transactions on Information Forensics and Security · doi:10.1109/tifs.2023.3265506 · Cited by 162
“The possible information leakage in the existing machine unlearning methods as discussed in [23] also motivated us to look towards a completely data-free approach to unlearning.”
Xu, H., Zhu, T., Zhang, L., Zhou, W., & Yu, P. S. (2024) · ACM Computing Surveys · doi:10.1145/3603620 · Cited by 164
“Thus, a membership inference attack or a poisoning attack would reveal a signiicant amount of detailed information about the unlearned samples [78, 108].”
BackgroundUnrolling SGD: Understanding Factors Influencing Machine Unlearning
Thudi, A., Deza, G., Chandrasekaran, V., & Papernot, N. (2022) · 2022 IEEE 7th European Symposium on Security and Privacy (EuroS&P) · doi:10.1109/eurosp53844.2022.00027 · Cited by 135
“Machine unlearning has since been extended to deep neural networks (DNNs) [7, 8, 9, 10, 11, 12, 13], and its privacy implications have been studied [14, 15, 16].”
Background
Metadata
- Authors
- Chen, M., Zhang, Z., Wang, T., Backes, M., Humbert, M., & Zhang, Y.
- Affiliations
- Helmholtz Center for Information Security (DE)
- University of Virginia (US)
- CISPA Helmholtz Center for Information Security, Saarbrücken, Germany
- University of Lausanne (CH)
- Venue
- Proceedings of the 2021 ACM SIGSAC Conference on Computer and Communications Security
- Publisher
- ACM
- Published
- 2021-11-12
- Pages
- 896-911
- ISSN
- 1543-7221 (ISSN-L)
- OpenAlex
- W3023153742
- Semantic Scholar
- 4ebdc89a4dac0d05d8a2567e3b7849be198cb0ba
- Cited by
- 195
- References
- 66
- Citations by year
- 2026: 33 · 2025: 61 · 2024: 45 · 2023: 30 · 2022: 20 · 2021: 6
- FWCI
- 11.67
- Type
- conference paper
- Language
- en
- Open access
- green
- serval.unil.ch
- Locations
- ACM SIGSAC Conference on Computer and Communications Security (CCS) (publishedVersion)
- VU Research Portal (publishedVersion)
- SERVAL (Université de Lausanne) (publishedVersion, other-oa)
- arXiv (Cornell University) (submittedVersion)
- IRIS (submittedVersion)
- VU Research Portal (submittedVersion)
- arXiv (Cornell University)
- Funders
- Helmholtz Association (ZT-I-OO1 4)
- National Science Foundation (1931443)
- Purdue University (Bilsland Dissertation Fellowship)
- David and Lucile Packard Foundation (Packard Fellowship)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Privacy-Preserving Technologies in Data
- Physical Sciences › Computer Science › Artificial Intelligence › Adversarial Robustness in Machine Learning
- Social Sciences › Sociology and Political Science › Privacy, Security, and Data Protection
- Keywords
- machine unlearning, right to be forgotten, membership inference attacks, information leakage, temperature scaling, differential privacy, privacy preservation
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-11, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-11 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: When Machine Unlearning Jeopardizes Privacy. https://app.cochat.ai/papers/10.1145/3460120.3484756
- The paper
- Chen, M., Zhang, Z., Wang, T., Backes, M., Humbert, M., & Zhang, Y. (2021). When Machine Unlearning Jeopardizes Privacy. Proceedings of the 2021 ACM SIGSAC Conference on Computer and Communications Security. https://doi.org/10.1145/3460120.3484756
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.