Privacy, Provenance, and Unlearning
Privacy, unlearning, and provenance answer three different questions. Training-data privacy asks what an observer can learn about the people and records used to train a system. Machine unlearning asks whether the influence of specified training data can be removed after training. Provenance asks what evidence accompanies an output about its origin and edit history. None of these promises follows from the others. A model can protect training records without attaching provenance, and a signed output can still disclose private information.
That separation matters because each promise needs a different test. Privacy needs a protected unit and an attacker model. Unlearning needs a retained-data retraining reference. Provenance needs a named mechanism, signer or key, and a verification policy. A refusal, an edited answer, or a badge may be useful, but none establishes all three.
Four events, not one leak
The word memorization often stands in for several different events:
- Memorization is a measured dependence between a trained model and a training example. Its meaning depends on the metric, such as exposure or likelihood relative to a reference model.
- Extractability means that an attacker can recover an unknown training sequence under a stated prompt, decoding strategy, access level, and query budget. A memorized sequence is not necessarily extractable.
- Disclosure means that released output reveals sensitive information about a person. An extracted sequence is not necessarily personal data, and disclosure can also come from a prompt, retrieval result, public source, or inference rather than memorization.
- Membership inference asks whether a supplied candidate belonged to a training dataset. Nothing must be reproduced; dataset membership can itself be sensitive.
These are different events. Carlini et al. recovered verbatim examples, including personally identifying information, from GPT-2 through black-box queries (Carlini et al. 2021). Later measurements found that extraction and memorization varied with repetition, capacity, and available context in the models studied (Carlini et al. 2023). Those results demonstrate attacks, not an assurance that every model or record can be extracted. Membership inference is similarly conditional: the original shadow-model attack showed that black-box outputs can reveal membership (Shokri et al. 2017), while later work found that distribution shift can make pretraining attacks appear more effective than they are (Duan et al. 2024).
Write the privacy threat model first
A privacy result is uninterpretable until its scope is recorded:
- Privacy unit. Is the protected unit one token, example, document, account, or person? Record-level protection and user-level protection are different when one user contributes many records.
- Adjacency. Which adjacent datasets count as one person's presence or absence? State whether one unit is added, removed, or replaced.
- Attacker access. Record model version, prompts, output text, scores or log probabilities, auxiliary information, adaptive queries, and query budget.
- Success criterion. Define exact or approximate recovery, membership classification, false-positive constraints, and the harm to the affected person.
The same model can pass one threat model and fail another. Restricting log probabilities may frustrate one membership attack but leave text extraction possible. A low average attack score can also conceal a small group of highly exposed examples. Report tail behavior and results by relevant language and record type, not only a global mean.
Reduce risk before and after training
Start with data minimization and purpose limitation: do not collect or retain a record merely because it might be useful later. The lineage practices in Chapter 6 should cover source data, derived datasets, tokenized copies, and fine-tuning data. They should also distinguish deletion from pseudonymization; replacing a name with a reversible identifier does not remove the underlying person's data.
Exact, near-duplicate, and repeated-span detection address different forms of duplication. In one set of language-model experiments, deduplication reduced measured memorization and improved evaluation hygiene (Lee et al. 2022). Related work measured substantially lower extraction risk after deduplicating the studied corpora (Kandpal et al. 2022). Deduplication reduced measured memorization in those settings; it does not provide differential privacy, catch every paraphrase, or remove a fact repeated across independent sources.
PII detection has false positives and false negatives. Measure precision and recall by identifier type, format, language, and document source. Names are not the only risk: combinations of occupation, location, and dates can identify a person. Filtering also has to reach derived datasets, tokenized copies, and fine-tuning data rather than only the first raw-data snapshot.
Serving-time controls still matter. Access control, rate limits, query and output-length limits, restricted score access, output filtering, and monitoring can reduce exposure or raise the cost of an extraction campaign. Buffer a high-risk output until a privacy gate has decided whether to release it, as in Chapter 57. These controls do not remove training influence, cannot recall information already disclosed, and should not be described as unlearning. Protecting prompts, retrieved context, and request logs is a separate confidential-inference problem covered by Chapter 60.
The formal promise of differential privacy
In this formula, denotes a randomized training mechanism. For adjacent datasets and and every measurable set of possible outputs , let represent the first output probability:
Likewise, let represent the probability for the adjacent dataset:
The mechanism is -differentially private when the privacy-loss bound is
Here the privacy unit determines what adjacency means. The parameter is the multiplicative privacy-loss bound, so smaller is stronger when the rest of the mechanism is fixed. permits a bounded probability of exceeding that multiplicative guarantee. The definition compares distributions of possible trained artifacts; it does not say that two training runs produce identical weights or that the model can never output personal information (Dwork and Roth 2014).
DP-SGD applies the definition to iterative optimization (Abadi et al. 2016). For a sampled batch of size , let be the per-example gradient at step . Clip each gradient to clipping norm :
Then average the clipped gradients with Gaussian noise:
Here, is the noise multiplier and is the identity covariance shape. With learning rate , the optimizer makes the update
The batch size, sampling rate, number of steps, clipping norm, and noise multiplier all affect privacy and learning. Repeated access to private data causes composition: privacy loss accumulates across steps and other releases. A privacy accountant converts the sampling and noise schedule into the reported bound under stated assumptions. Every tuning run or private-data-dependent release must be included in that accounting.
Differential privacy bounds one unit's influence on the mechanism's output distribution. It does not remove information repeated across many privacy units, protect private inference prompts, or guarantee useful accuracy. Utility must be measured for the deployment population and rare cases, not inferred from the privacy budget alone.
Define unlearning against retraining
machine unlearning, a process that aims to remove the influence of specified training data from a trained model, needs a counterfactual target. Here denotes a randomized training algorithm, the original training dataset, the forget set, and the retained dataset. The symbol denotes an unlearning algorithm applied to an artifact produced by . The retraining reference is
Exact unlearning makes the distribution of artifacts on the left match the distribution produced by retained-data training on the right. Approximate unlearning bounds a stated distance or distinguishability between those distributions (Guo et al. 2020). The target is a distribution because training and unlearning can both contain randomness. Matching one checkpoint or one answer is not the definition.
Full retained-data retraining is the general reference, not the only possible exact method. Some restricted learners support efficient exact or certified removal under their mathematical assumptions. Those results do not transfer automatically to a different architecture, optimizer, or training pipeline. Even exact unlearning removes the specified records' influence under the declared training algorithm; it need not erase a semantic fact also present in retained duplicates, correlated examples, or earlier pretraining.
SISA: deletion designed into training
SISA partitions the training data into disjoint shards, divides each shard into cumulative slices, trains isolated component models, and saves intermediate states. A deletion can retrain only affected components from a checkpoint before the relevant slice, then repeats the declared aggregation (Bourtoule et al. 2021). The training procedure must be designed in advance, and exactness is relative to that procedure.
The assumptions are operational as well as mathematical. Shard and slice assignments, randomness, checkpoints, and aggregation must be reproducible. Multiple affected shards require multiple retrains, and a learned aggregator may itself need deletion. The original evidence concerned supervised classification; it does not make arbitrary post-hoc editing exact or establish the same trade-off for a frontier language model.
Evaluate forgetting, retention, and recovery
An empirical evaluation should compare several unlearned artifacts with several retained-only retraining references. At minimum, report:
- Forget-set behavior: likelihood, exposure, extraction attempts, membership inference, direct prompts, paraphrases, and adaptive prompts.
- Retain-set utility: task performance, calibration, and behavior on neighboring knowledge. Measure collateral damage rather than hiding it in an average.
- Recovery: white-box probes where available, short fine-tuning or relearning attacks, and access to pre-deletion outputs or artifacts.
- Operations: deletion size, sequential requests, storage, latency, cost, and the complete affected-artifact inventory.
One refusal is not evidence of unlearning. A method may suppress the tested answer while retaining membership or enabling rapid relearning. Conversely, a retained-data retrain may still state the same fact because retained sources support it. TOFU provides synthetic author profiles, forget sets, retained references, and multiple metrics; its tested baselines did not jointly achieve effective forgetting and retention (Maini et al. 2024). It is a useful benchmark, not a universal certificate.
Deletion is a lineage operation
A deletion request should resolve from the data subject or record scope to every affected artifact:
- the source record, dataset snapshots, deduplicated copies, transformed and tokenized data, and queued future training runs;
- checkpoints, optimizer state, adapters, merged models, ensembles, and model registry copies;
- embeddings, retrieval indexes, evaluation fixtures, caches, logs, and backups governed by the same request; and
- a deletion tombstone that prevents accidental re-ingestion while preserving the minimum evidence needed for audit.
A checkpoint and adapter copy left outside the inventory can restore the old behavior after the primary model is replaced. The tombstone needs a stable scope identifier and expiry or review policy, without retaining the deleted content itself.
Editing is a different operation
Knowledge editing is not machine unlearning. It changes selected model behavior, often a subject-relation answer, and is commonly evaluated for direct efficacy, paraphrase generalization, and locality (Meng et al. 2022; Meng et al. 2023). It does not establish that a training record's influence is gone. A successful edit may fail on a paraphrase or logical consequence, cause collateral changes, or leave membership evidence intact. Use editing when the product promise is to change an answer; do not report it as record deletion without a corresponding unlearning argument and evaluation. This distinction also limits what mechanistic findings from Chapter 54 can establish.
Provenance provides evidence, not truth
A statistical watermark, a signed manifest, attribution to an author, and the truth of the content are four separate claims. A participating generator can embed a detectable signal without identifying the person who requested the output. A signer can attest to an edit history without proving that the depicted event happened.
Statistical watermarks
One family of text watermarks changes generation so tokens correlate with a secret keyed pattern. A detector computes a score and compares it with a detector threshold (Kirchenbauer et al. 2023). The operational report needs the false-positive rate, false-negative rate, abstention policy, text length, language, sampling configuration, model version, and measured quality. It also needs the evaluation base rate: even a small false-positive rate can dominate when watermarked text is rare.
Robustness is transformation-specific. Test truncation, ordinary edits, watermark paraphrase, translation, model rewriting, and mixed human/model text. A detector may retain confidence after one edit and lose it after another. Absence of a detectable watermark is not proof of human authorship; the text may come from a nonparticipating generator, be too short, or have been changed.
SynthID-Text is a production example with a different sampling construction from the original green-list scheme. Its large-scale evaluation covered about twenty million Gemini responses and measured detection and quality in that deployment (Dathathri et al. 2024). This evidence supports the tested mechanism and configuration, not a universal AI-text detector.
Signed Content Credentials
C2PA represents provenance claims in a signed manifest. The manifest contains assertions, an asset hash or other content binding, a claim, a claim signature, and information needed to evaluate the signing credential (Coalition for Content Provenance and Authenticity 2026). A valid signature and binding provide tamper evidence: they show that the signed claim was bound to this asset and has not been modified undetected since signing. Linked manifests can record an edit history when participating tools preserve or extend the chain.
Verification still requires policy. The verifier decides which trust anchors, credential status, algorithms, and assertion types it accepts. An unknown signer is not automatically malicious, but its claims may be outside the verifier's trust policy. A valid credential does not prove that the depicted event happened, that every edit was recorded, or that the assertions were honest.
Manifests can be embedded or stored externally, and soft bindings can help recover provenance after a permitted transformation. Complete removal remains possible. A stripped manifest produces missing evidence, not a successful verification; an unsigned asset is not evidence that the asset is synthetic. Watermarks and credentials can therefore complement one another, but neither turns absence into proof. C2PA's manifest architecture is also the subject of an ISO work item (International Organization for Standardization 2026). Current standards and disclosure duties are tracked in Chapter 61 because their status and scope change independently of these engineering guarantees.
Keep an operating record
The audit record should be machine-readable and versioned. A compact schema can use these fields:
data_subject_or_record_scope:
source_and_legal_basis:
dataset_and_transform_revisions:
privacy_unit_and_adjacency:
dp_mechanism_and_accountant:
training_job_and_artifact_revisions:
deletion_request_and_tombstone:
affected_artifact_inventory:
unlearning_method_and_retraining_reference:
forget_retain_and_privacy_results:
provenance_mechanism_and_signer:
detector_threshold_and_error_rates:
verification_and_exception_owner:
Do not let the schema become a claim generator. Empty fields, exceptions, and failed verification must remain visible. Access to the record should be restricted because lineage and deletion evidence may themselves identify a person.
Regression scenarios
Automate cases that exercise the assumptions, not only the happy path:
- a duplicated canary, one user across many records, a PII detector miss, and membership distribution shift;
- a checkpoint and adapter copy, a forget-set paraphrase, a relearning attack, and a retained-utility regression;
- a watermark paraphrase, a human-text false positive, a stripped C2PA manifest, and a tampered asset.
Each failure should name the affected person or content claim, the responsible owner, and whether the remedy changes data, training, serving, deletion, or verification.
- What empirical evidence is enough for approximate unlearning? Formal definitions compare distributions, but practical language-model evaluations use finite attacks and utility tests. Passing today's probes does not rule out a stronger adaptive or white-box test tomorrow.
- Which watermark threat model is useful? A watermark can work well for aggregate measurement under known configuration while remaining vulnerable to deliberate transformation. Product claims should match the tested editor, text length, language, and error rate.
- Where is differential privacy worth its utility cost? The answer depends on the privacy unit, data regime, model, budget, and acceptable error. The guarantee and the utility measurement should be reported together.
Training and data systems determine which privacy and deletion promises are available later. Complete lineage enables scoped deletion; a declared privacy unit and accountant enable a differential-privacy claim; reproducible shards, slices, randomness, and aggregation enable SISA-style retraining. Serving then controls what is released, while the provenance path preserves evidence about an output. These layers constrain one another, but none substitutes for the others. When the lower layer cannot identify affected data or reproduce the training procedure, the upper layer cannot manufacture that evidence after the fact.
Three promises, three kinds of evidence
Privacy is supported by a threat model, measured attacks, data controls, and, where used, a formal privacy budget. Unlearning is supported by lineage and a comparison with retained-data retraining. Provenance is supported by a watermark detector or verified signed assertions under a trust policy. Keeping these claims separate makes them more useful: an operator can state what was protected, what was removed, what was signed, and what remains unknown.
Further reading
- Carlini et al., “Extracting Training Data from Large Language Models” (verbatim extraction of PII from a deployed model), 2021. arXiv:2012.07805Carlini et al. show that large language models memorize training data, and that black-box query access to GPT-2 can extract verbatim text including personally identifiable information.
- Carlini et al., “Quantifying Memorization Across Neural Language Models” (memorization scales with model size, duplication, and context), 2023. arXiv:2202.07646Carlini et al. quantify three log-linear relationships showing that LLM memorization of training data grows with model capacity, data duplication, and context length, and is more prevalent than previously believed.
- Shokri et al., “Membership Inference Attacks Against Machine Learning Models” (the generalization-gap attack), 2017. arXiv:1610.05820This paper introduces membership inference attacks that use shadow training to determine, via black-box API access, whether a record was in a model's training dataset.
- Duan et al., “Do Membership Inference Attacks Work on Large Language Models?” (distribution shift can confound pretraining membership inference), 2024. arXiv:2402.07841Duan et al. find that tested membership-inference attacks on large language model pretraining data often perform near chance and show how distribution shift can create misleading results.
- Kandpal et al., “Deduplicating Training Data Mitigates Privacy Risks in Language Models” (measured links among duplication, memorization, and extraction), 2022. proceedings.mlr.pressKandpal et al. measure a superlinear relationship between duplication and memorization and substantially lower extraction risk after deduplicating the studied language-model corpora.
- Dwork, Cynthia; Roth, Aaron. The Algorithmic Foundations of Differential Privacy (formal definitions, composition, and mechanisms for differential privacy). Now Publishers, 2014. doi.orgDwork and Roth develop the formal foundations of differential privacy, including adjacency, privacy parameters, composition, and standard mechanisms.
- Abadi et al., “Deep Learning with Differential Privacy” (DP-SGD: clipping, noise, privacy accounting), 2016. arXiv:1607.00133Abadi et al. introduce DP-SGD, training deep neural networks under differential privacy via per-example gradient clipping, Gaussian noise, and a moments accountant for tighter privacy budget tracking.
- Guo et al., “Certified Data Removal from Machine Learning Models” (certified removal defined against retained-data retraining), 2020. proceedings.mlr.pressGuo et al. define certified removal by indistinguishability from retained-data retraining and construct a mechanism for regularized linear models under stated assumptions.
- Bourtoule et al., “Machine Unlearning” (the SISA sharded-retraining scheme), 2021. arXiv:1912.03817This paper introduces SISA training, which partitions data into shards and slices to reduce the retraining cost of machine unlearning, achieving up to 4.63x speedup over full retraining.
- Maini et al., “TOFU: A Task of Fictitious Unlearning for LLMs” (synthetic forget sets and retained-data references for LLM unlearning), 2024. arXiv:2401.06121TOFU evaluates language-model unlearning on synthetic author profiles with forget sets, retain sets, and multiple behavioral and utility metrics.
- Meng et al., “Locating and Editing Factual Associations in GPT” (ROME), 2022. arXiv:2202.05262ROME uses causal mediation analysis to locate factual associations in mid-layer feed-forward modules of GPT and introduces a rank-one weight editing method to update specific facts.
- Meng et al., “Mass-Editing Memory in a Transformer” (MEMIT), 2023. arXiv:2210.07229MEMIT scales knowledge editing in large language models to thousands of simultaneous fact updates by distributing parameter changes across a range of critical MLP layers.
- Kirchenbauer et al., “A Watermark for Large Language Models” (green-list sampling watermark), 2023. arXiv:2301.10226This paper proposes a statistical watermarking framework for LLM output that embeds detectable signals into generated text by biasing token sampling toward a randomized "green list," detectable from as few as 25 tokens without model access.
- Coalition for Content Provenance and Authenticity, “C2PA Technical Specification” (Content Credentials for signed media provenance), 2026. c2pa.orgThe C2PA specification defines signed provenance manifests for recording a media asset's origin and edits without claiming that the content itself is true.
- Dathathri et al., “Scalable Watermarking for Identifying Large Language Model Outputs” (SynthID-Text: the sampling-bias watermark hardened for and deployed in production), 2024. nature.comSynthID-Text productionizes generation-time watermarking with negligible latency, was deployed on live Gemini traffic in a 20-million-response test, and is open-sourced.
- International Organization for Standardization, “ISO/CD 22144: Authenticity of Information, Content Credentials” (an ISO work item for content credentials), 2026. iso.orgThe draft ISO standard develops interoperability requirements for content credentials and authenticity metadata.
Comments
Log in to comment