The DOI matches its Crossref record; 32 of the 32 references it cites were found in a registry or an index.
Analyzing Information Leakage of Updates to Natural Language Models
Open accessAbstract
To continuously improve quality and reflect changes in data, machine learning applications have to regularly retrain and update their core models. We show that a differential analysis of language model snapshots before and after an update can reveal a surprising amount of detailed information about changes in the training data. We propose two new metrics---differential score and differential rank---for analyzing the leakage due to updates of natural language models. We perform leakage analysis using these metrics across models trained on several different datasets using different methods and configurations. We discuss the privacy implications of our findings, propose mitigation strategies and evaluate their effect.
Overview
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
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
“Recently, existing research has shown that the unlearning operation not only does not reduce the risk of user privacy leakage but actually increases this risk [106, 107].”
From Machine Unlearning: A Survey (2024) · cites this paper as [106]
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Training data extraction attacks have until now been limited to small LMs trained on academic datasets under artificial training setups (e.g., for more epochs than typical) [8, 66, 68, 73], or settings where the adversary has a priori knowledge of the secret they want to extract (e.g., a social security number) [8, 27].”
From Extracting Training Data from Large Language Models (2020) · cites this paper as [73]
Experimental — accuracy not yet measuredIn the citing paper
“Prior work shows that (100 × to 1000 × smaller) language models potentially memorize training data in semi-realistic settings [8, 73].”
From Extracting Training Data from Large Language Models (2020) · cites this paper as [73]
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Citations
Cites
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017) · arXiv · doi:10.48550/arxiv.1706.03762 · Cited by 120,404 · cited 2×
“The most prominent are Recurrent Neural Networks (RNNs) using Long Short-Term Memory [10] cells (or variants thereof) and the more recent Transformers [15, 21].”
MethodHochreiter, S., & Schmidhuber, J. (1997) · Neural Computation · doi:10.1162/neco.1997.9.8.1735 · Cited by 102,318 · cited 1×
“The most prominent are Recurrent Neural Networks (RNNs) using Long Short-Term Memory [17] cells (or variants thereof) and the more recent Transformers [23, 30].”
MethodBERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019) · Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) · doi:10.18653/v1/n19-1423 · Cited by 54,325 · cited 2×
“(2) a model based on the Transformer architecture [21] (more concretely, using the BERT [6] codebase) with four layers of six attention heads, each with a hidden dimension of 192.”
BackgroundMethodBuilding a Large Annotated Corpus of English: The Penn Treebank
Marcus, M. (1993) · doi:10.21236/ada273556 · Cited by 7,405
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 · cited 1×
“At a high level, differential privacy can be achieved in gradientbased optimization computations [1, 3, 20] by clipping the gradient of every record in a batch according to some bound L, then adding noise proportional to L to the sum of the clipped gradients, averaging over the batch size and using this noisy average…”
BackgroundThe 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 · cited 2×
“Recall that an ε-DP algorithm F is kε-differentially private for groups of size k [8].”
BackgroundMembership 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 2×
“data (unlike [18]) and the auditor only observes a list of several top-ranked tokens.”
BackgroundThe Algorithmic Foundations of Differential Privacy
Dwork, C., & Roth, A. (2013) · doi:10.1561/9781601988195 · Cited by 3,856
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 · cited 1×
“…reconstruction attacks [24] against the updated model, as the The goal of membership inference attacks [25, 26] is weaker in Finally, model inversion attacks [12, 13] repurpose a model to work backwards , inferring unknown attributes of individuals given Relationship to differential privacy.…”
BackgroundAdvances in Neural Information Processing Systems 32 (NeurIPS 2019)
Wallach, H., Larochelle, H., Beygelzimer, A., d’Alché–Buc, F., Fox, E. B., & Garnett, R. (2020) · HAL · Cited by 2,451
Research works printed but not linked: 4
- Bourtoule Lucas, 2021, Machine Unlearning. In 42nd IEEE Symposium on Security and Privacy, S&P
- Dwork Cynthia, 2010, ICS
- Guo Chuan, 2020, 37th International Conference on Machine Learning, ICML
- Vaswani Ashish, 2017, NIPS
Cited by
Extracting Training Data from Large Language Models
Carlini, N., Tramer, F., Wallace, E., Jagielski, M., Herbert-Voss, A., Lee, K., Roberts, A., Brown, T., Song, D., Erlingsson, U., et al. (2020) · arXiv · doi:10.48550/arxiv.2012.07805 · Cited by 727
“Prior work shows that (100 × to 1000 × smaller) language models potentially memorize training data in semi-realistic settings [8,73].”
BackgroundSecurity and Privacy Challenges of Large Language Models: A Survey
Das, B. C., Amini, M. H., & Wu, Y. (2025) · ACM Computing Surveys · doi:10.1145/3712001 · Cited by 445
Machine Unlearning of Features and Labels
Warnecke, A., Pirch, L., Wressnegger, C., & Rieck, K. (2023) · Network and Distributed System Security Symposium · doi:10.14722/ndss.2023.23087 · Cited by 381
Analyzing Leakage of Personally Identifiable Information in Language Models
Lukas, N., Salem, A., Sim, R., Tople, S., Wutschitz, L., & Zanella-Béguelin, S. (2023) · IEEE Symposium on Security and Privacy (S&P) · doi:10.1109/sp46215.2023.10179300 · Cited by 552
“Research has shown the feasibility of extracting different types of information including individual sentences [9, 28], inserted canaries [8, 50, 68] as well as n -grams [42].”
BackgroundSelf-citationAdversarial machine learning : a taxonomy and terminology of attacks and mitigations
Vassilev, A., Oprea, A., Fordyce, A. J., & Anderson, H. S. (2024) · doi:10.6028/nist.ai.100-2e2023 · Cited by 317
Metadata
- Authors
- Zanella-Béguelin, S., Wutschitz, L., Tople, S., Rühle, V., Paverd, A., Ohrimenko, O., Köpf, B., & Brockschmidt, M.
- Affiliations
- Microsoft Research (United Kingdom) (GB)
- The University of Melbourne (AU)
- ORCID
- Venue
- Proceedings of the 2020 ACM SIGSAC Conference on Computer and Communications Security
- Publisher
- ACM
- Published
- 2020-10-30
- Pages
- 363-375
- ISSN
- 1543-7221 (ISSN-L)
- OpenAlex
- W3027379683
- Semantic Scholar
- 782e7fba4dfa8e2095a71cfcf85995543f37ac75
- Cited by
- 86
- References
- 32
- Citations by year
- 2026: 4 · 2025: 18 · 2024: 17 · 2023: 22 · 2022: 15 · 2021: 7 · 2020: 3
- FWCI
- 4.48
- Type
- conference paper
- Language
- en
- Open access
- green
- arxiv.org
- Locations
- ACM SIGSAC Conference on Computer and Communications Security (CCS) (publishedVersion)
- arXiv (Cornell University) (submittedVersion)
- arXiv (Cornell University)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Privacy-Preserving Technologies in Data
- Physical Sciences › Computer Science › Artificial Intelligence › Adversarial Robustness in Machine Learning
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Keywords
- information leakage, natural language models, model updating, differential analysis, privacy, retraining, privacy mitigation
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.
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: Analyzing Information Leakage of Updates to Natural Language Models. https://app.cochat.ai/papers/10.1145/3372297.3417880
- The paper
- Zanella-Béguelin, S., Wutschitz, L., Tople, S., Rühle, V., Paverd, A., Ohrimenko, O., Köpf, B., & Brockschmidt, M. (2020). Analyzing Information Leakage of Updates to Natural Language Models. Proceedings of the 2020 ACM SIGSAC Conference on Computer and Communications Security. https://doi.org/10.1145/3372297.3417880
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.