RESEARCH DOSSIER
Privacy-Preserving Statistical Learning for Insurance
Abstract. Insurance statistics converts detailed and often sensitive information about persons, contracts, exposures, illnesses, property, injuries, claims, and settlements into predictive estimates and financial decisions. A conventional data-consolidation architecture makes these records available to analysts and learning systems; a privacy-preserving architecture must constrain what an adversary can infer from collection, intermediate computation, model parameters, released statistics, and repeated model queries. This monograph establishes the mathematical and institutional foundations of privacy-preserving statistical learning for insurance. It examines record linkage, membership and gradient inference, the meaning of the protected person and the neighboring dataset, differential privacy and sensitivity, local and central randomization, private statistical estimation, differentially private learning, federated optimization, secure aggregation, and their compositions. It evaluates disclosure control jointly with actuarial credibility, class imbalance, claims-tail inference, fairness, calibration, temporal drift, operational security, and jurisdiction-dependent obligations. Explicit hypothetical calculations show how privacy noise can be negligible in a large count yet overwhelm a rare insurance subgroup, and how truncating severity contributions changes the estimand as well as the disclosure bound. The analysis distinguishes formal privacy guarantees from encryption, access control, data minimization, and legal compliance. An original analytical synthesis identifies the insured-person contribution boundary, target estimand, release surface, composition ledger, and conditional accuracy as jointly necessary specifications for interpreting a claimed privacy-preserving insurance model. The synthesis is a critical application of existing mathematical results, not a proposed new formal theory. Original insurer datasets and external empirical evaluations are not claimed. All regulatory discussion identifies the instrument and relevant jurisdiction, and the bibliography separates original scientific results from institutional guidance.
Keywords: Insurance statistics; statistical disclosure control; differential privacy; personal data; actuarial modeling; claims data; membership inference; gradient leakage; epsilon; delta; sensitivity; bounded contribution; insurance frequency; claim severity; risk classification; private estimation; differential privacy stochastic gradient descent; federated learning; secure aggregation; secure multiparty computation; privacy accounting; privacy-utility frontier; rare events; calibration; fairness; exposure; claims reserving; data governance; GDPR; EIOPA; NAIC; auditability.
Introduction: Research Object, Evidence, and Method
Insurance learning begins with contractual and statistical questions. A frequency model estimates the expected number of claims for a defined unit of earned exposure. A severity model estimates the distribution of individual payments subject to coverage and settlement processes. A reserve model predicts cash flows for claims already incurred. Mortality, morbidity, lapse, fraud, and catastrophe models target different outcomes, time windows, and populations. Each task requires data that may be unusually revealing: a rare diagnosis, a highly distinctive loss, a household address, a claim investigation note, or an accident time may permit identification even after names have been deleted. The scientific question is how these observations can inform reproducible inference while constraining additional disclosure attributable to the learning operation.
The foundational differential privacy analysis of Dwork (2006) characterizes protection through the small influence of the inclusion or exclusion of one participant. Dwork and Roth (2014), in The Algorithmic Foundations of Differential Privacy, systematize randomization, composition, and accuracy. The 2025 NIST evaluation guidelines emphasize that an asserted mathematical guarantee must be evaluated together with implementation choices, privacy units, and actual deployment conditions. These sources establish the analytical departure point. No single parameter, software library, or cryptographic primitive by itself defines a complete privacy practice.
The study uses four methods. First, it states data-generating and observation assumptions explicitly and derives mathematical consequences for statistical queries and prediction. Second, it compares relevant original scientific literature, with claims limited to the accessible publisher records, source texts, and institutional materials examined. Third, it constructs hypothetical arithmetic examples whose figures can be independently recomputed. Fourth, it compares EU statutory provisions, EU insurance supervisory guidance, US state-oriented model guidance, and US actuarial standards without treating different instruments as interchangeable. The analytical claims are conditional on the stated threat and population definitions. Hypothetical values are educational calculations rather than observed insurer results.
A central distinction concerns the output being protected. A protected training dataset may produce a published table, a model accessible through an application programming interface, a score delivered to an insurer, or an aggregate transmitted to another institution. Identical methods offer different practical protection when these outputs, their access histories, and the adversary's auxiliary information differ. For a risk classifier, the strongest guarantee is meaningful only relative to its neighboring datasets; for a mortality estimate, the perturbation's utility is meaningful only relative to its exposure denominator, group composition, and decision horizon. Privacy and inferential validity must be evaluated against the same real research unit.
Defined terminology and notation
Let D be the available insurance database, U a statistical unit such as a person, H the hypothesis or model class, A a randomized training or query algorithm, and M(D) the released result. Let theta denote the population estimand and T(D) the nonprivate estimator. Privacy loss is defined with respect to a declared neighboring-dataset relation, D ~ D', where the datasets differ in the information associated with one protected unit. The privacy parameters epsilon and delta describe a bound on output-distribution differences; they are not empirical probabilities of reidentification. Exposure E is earned person-years, policy-years, vehicle-years, or another contractually meaningful denominator. Claims C and aggregate paid or incurred loss S may be measured on radically different scales. All monetary examples below use hypothetical US dollars solely for dimensional clarity.
For multiple policies held by one person, a one-policy definition of the protected unit is materially weaker than a one-person definition. For several insured persons named on a contract, even the person-to-policy relation requires explicit handling. An insurer must decide which records may move together under the adjacency rule before choosing epsilon, calculating global sensitivity, or making a public privacy statement.
Insurance Personal Data and Risk
Observations, persons, and insurance contracts
An insurance dataset is often a longitudinal administrative record rather than a random sample drawn once from a stable population. An application is followed by underwriting decisions, premium changes, coverage alterations, claims occurrence, reporting delays, adjustments, reserves, payments, recovery, cancellation, and renewal. The probability that an observation enters an analytic dataset depends on contractual eligibility, marketing, prior claims, selection, reporting incentives, and operational systems. A policyholder can have several policies; a claim can involve insureds, claimants, beneficiaries, injured third parties, and investigators. A single event can create multiple ledger rows. The row count therefore has no universal correspondence to persons, independent events, or units of disclosure risk.
The EU General Data Protection Regulation (European Parliament and Council, 2016) defines personal-data principles, including purpose limitation, data minimization, accuracy, integrity, confidentiality, and accountability in Article 5. Article 9 addresses special categories of data, Article 25 concerns data protection by design and default, and Article 32 concerns security of processing. Article 22 addresses specified automated individual decision-making, subject to its conditions and exceptions. These provisions apply in their legal scope and should not be read as a statistical definition of anonymization. A controller must separately establish an applicable legal basis, any required exception for sensitive information, lawful purposes, retention arrangements, and rights handling. A formal DP guarantee can support risk management while leaving those legal questions to the appropriate jurisdictional analysis.
Consider a hypothetical health-insurance file containing diagnosis, age at treatment, district, procedure, incurred amount, and dates. Removing policy numbers and names does little against an adversary who knows of a unique hospital episode. A commercial insurance portfolio may contain large industrial loss amounts identifiable from public news reports. Automobile claims may embed precise coordinates, driving trajectories, or photographs. The adversary's auxiliary knowledge is not limited to information the data holder was willing to disclose. Privacy engineering must therefore treat predictable linkage pathways, released outputs, and external data as a joint disclosure setting.
Definition of the target population
Suppose theta is the next-year claim frequency among a specified insured cohort. The analyst observes N claims over E policy-years. The naive estimate N/E is a valid empirical ratio for the recorded exposure; its interpretation as a prospective probability or intensity depends on complete exposure, a stable time basis, correct claim definitions, and the representativeness of the insured population. Differential privacy can randomize N, but it cannot remove market selection, missing claims, or a change in product design. A privacy-preserving model may be precisely private and systematically biased for the commercial target. This is why the estimand and the observation mechanism belong in a scientific privacy assessment.
Statistical data quality is an actuarial requirement independent of the security mechanism. The US Actuarial Standards Board's ASOP No. 23, Data Quality (revised 2016) discusses selection, review, use, and limitations of data for actuarial work. The companion ASOP No. 56, Modeling (revised 2019, effective October 2020) addresses model risk, implementation, validation, and use under its stated US professional scope. Neither document provides a general-purpose certificate that an algorithm is private. They illustrate why governance must connect data adequacy and model validation to protection of individual records.
Privacy units and contribution bounds
The insured person is an intuitive protection unit when modeling an individual's exposure, but household effects may justify a family or address-level unit in some settings. If a person contributes k claims, adjacency defined as the removal of a single claim does not control the full effect of removing that person's history. A claim-level guarantee and a person-level guarantee answer different questions. Group privacy can sometimes translate a record-level result into a weaker bound for several record changes; operational systems should instead bound a person's total contribution when the statement promises person-level protection.
Contributions can be constrained by retaining at most a specified number of records per person, capping per-person sums, clipping a per-person gradient, or sampling persons with specified rules. Each choice changes the statistical signal. Claims severity is especially sensitive because outliers may dominate aggregate losses and capital measures. An uncapped monetary sum can have very high or unbounded global sensitivity under the model's data domain; consequently, the theoretical variance of calibrated noise can become unacceptable. Capping produces a finite bound, but then the estimator targets a capped-loss quantity. The expected uncovered tail contribution must be estimated or acknowledged separately, and any tail correction trained on the same sensitive data requires its own privacy accounting.
Operational privacy boundaries
Data minimization reduces the number of attributes exposed; access controls restrict which actors can obtain them; encryption protects information against specified interception or storage threats; secure computation constrains intermediate visibility; formal output privacy bounds the influence of participation on released computations. A robust design may employ all five. Their protection claims should remain distinct because a breach of one surface need not imply compromise of every other surface, and a mathematically private release does not protect files held elsewhere by an authorized but compromised user. The NIST Privacy Framework provides voluntary organizational privacy-risk management guidance, whereas statistical mechanisms specify formal properties of defined computational outputs.
Statistical Disclosure and Reidentification
Disclosure as an inference problem
An adversary may attempt identity attribution, attribute inference, membership inference, reconstruction, or model inversion. Identity attribution joins information to a real person; attribute inference estimates a private characteristic given another set of facts; membership inference asks whether a person's record helped train a model; reconstruction aims to recover particular records or features from a released output. An attack's success depends on priors, access, side information, model confidence, sampling, and the loss function used to judge success. The fact that a particular test attack failed does not establish privacy for all plausible adversaries.
Shokri, Stronati, Song, and Shmatikov (2017), in their membership inference experiments, show that black-box model predictions can disclose training-set participation in realistic classification tasks. Their experiments support a general vulnerability mechanism, not a numerical leakage rate transferable to every insurer, model architecture, or portfolio. A policyholder's inclusion in a dataset may itself reveal a specific event, especially where training was restricted to specialized claims or treatment episodes.
The gradient reconstruction research of Zhu, Liu, and Han (2019) demonstrates that intermediate gradients can carry information allowing recovery of aspects of training data under its experimental conditions. These attacks are material to distributed or federated learning where institutional actors might exchange gradients or updates. A system that never transmits raw claims still exposes an attack surface if it releases small-batch updates, per-client gradients, unusually detailed confidence scores, or adaptive queries.
Why aggregation alone is insufficient
Suppose a public table gives the total claim amount for a geographic cohort and a separate update gives the same total after one insured leaves. The difference may reveal that insured's claim amount even though neither released table contains a name. Overlapping subtotals, repeated annual reports, and interaction with outside information can create related attacks. When aggregate values are exact, their sensitivity to participation may be transparent to the adversary. Security controls on the training source do not resolve what is mathematically inferable from final outputs.
The precise risk cannot be assessed from one output in isolation where a program releases many tables. Two individually innocuous summaries may jointly isolate a rare diagnosis or injury. A publication ledger should specify each requested statistic, user audience, neighboring dataset definition, noise mechanism, and cumulative privacy cost. Aggregate suppression and minimum-cell rules can reduce some risks, but their effectiveness relies on the total release pattern and available auxiliary records. Differential privacy offers a formal response to arbitrary side information under an explicit adjacency definition, provided the mechanism and its implementation match the proof.
Threat models for the insurance lifecycle
At data ingestion, the threat includes external interception and an insider with access to originals. At training, it includes model developers, other collaborating institutions, cloud infrastructure, intermediate outputs, and software vulnerabilities. At release, it includes model consumers, business partners, brokers, public researchers, and regulators. At decision time, it includes a model score that may be observed only for one applicant and may reveal training patterns indirectly. A privacy threat model needs to state which of these parties is trusted with identifiable input and which is controlled by the adversary. A mechanism can be private against the prediction recipient while relying on a trusted training operator; this is a meaningful but bounded claim.
A separate scientific risk is population disclosure: learning the prevalence of an adverse event among a group may be socially sensitive even if a mechanism bounds any one person's marginal influence. Individual differential privacy does not remove every ethical concern over correlations, use of group statistics, or harmful decisions. Disparate economic outcomes may follow from accurate population estimates. Privacy, fairness, and validity therefore require their own tests and governance, while connections among them are examined in the same insurance use case.
Differential Privacy and Noise
Formal definition and adjacency
A randomized mechanism M satisfies (epsilon, delta)-differential privacy if, for every neighboring pair D ~ D' and every measurable output event S,
P[M(D) in S] <= exp(epsilon) P[M(D') in S] + delta.
When delta = 0 the guarantee is pure epsilon-differential privacy. The probabilities concern the mechanism's randomization conditional on possible data sets; epsilon and delta are constraints on distributional divergence rather than direct predictions of an individual's probability of discovery. The strength of the statement rests on the adjacency relation, data-preprocessing assumptions, random number generation, and all output channels counted as the release. A smaller epsilon generally means a tighter worst-case bound for a fixed adjacency rule, but there is no context-free universal safe epsilon. NIST's differential-privacy evaluation guidance (Near, Darais, Lefkovitz, and Howarth, 2025) treats design details and implementation pitfalls as decisive to assessment.
For a scalar statistic f, its global L1 sensitivity is Delta_1 f = sup_{D ~ D'} |f(D)-f(D')|. When one person contributes a single binary claim indicator, count sensitivity is one under add/remove-person adjacency. When one person may contribute ten claims, the unbounded total count no longer has sensitivity one unless claims are clipped or the data model restricts each person's total contribution. If the denominator is a public fixed exposure E, a released rate f(D)=N(D)/E has sensitivity 1/E for the single-indicator case. If exposure also changes with the person's participation, simple division by E is no longer a fixed-denominator query, and private numerator and denominator mechanisms or a justified public denominator are required.
Laplace mechanism and a reproducible count example
Dwork, McSherry, Nissim, and Smith (2006), in Calibrating Noise to Sensitivity, establish the use of sensitivity-calibrated randomization for private statistical queries. For a scalar query with finite Delta_1 f, the Laplace mechanism releases T=f(D)+Z, where Z has mean zero and scale b=Delta_1 f/epsilon. Its variance is 2b^2. This is a result conditional on the chosen adjacency, clipping, and correct distributional implementation, rather than a license to add arbitrary independent noise after disclosing the originals.
Take a hypothetical portfolio with a fixed, nonprivate denominator of E=10,000 person-years and 1,000 eligible binary claims, one contribution per person. At epsilon=0.5, the Laplace noise scale for the count is 2 and its standard deviation is sqrt(8), approximately 2.828 claims. For the rate N/E, the noise standard deviation is 2.828/10,000 = 0.0002828, equivalent to 0.02828 percentage points. The nonprivate observed rate is 0.1. The ratio between this illustrative privacy standard deviation and the observed rate is about 0.2828%. This comparison does not establish statistical adequacy: sampling error, data completeness, subgroup structure, and deployment decisions remain to be examined.
Now consider a rare cohort with E=50 person-years and one observed binary claim. Under the same unit definition and epsilon, the count-noise standard deviation remains approximately 2.828, but the rate-noise standard deviation increases to 2.828/50 = 0.05657, or 5.657 percentage points, against an observed frequency of 2%. The output may be negative or greater than one before any range restriction. Constraining the output to a plausible interval is a postprocessing operation, which can retain the original DP guarantee, but it introduces bias and changes distributional uncertainty. A private measurement that is workable for a full portfolio may be scientifically weak for an infrequent condition or a small geographic cell.
Bounded loss totals and changed estimands
Consider a hypothetical incurred-loss statistic S_B equal to the sum of per-person losses truncated to the range [0, B], with B=$100,000 and add/remove-person adjacency. Its global sensitivity is B. At epsilon=0.5, Laplace scale is B/epsilon=$200,000, with a noise standard deviation of approximately $282,843. If exactly 10,000 fixed public exposure units are used as a denominator, the noise standard deviation of the capped loss per exposure is approximately $28.2843. These are mechanism variance calculations; no confidential experience is modeled or simulated. When extreme underlying losses exceed B, S_B estimates the aggregate of capped contributions rather than the unrestricted aggregate. No unproven claim about an insurance loss tail can make the clipping bias disappear.
For example, a single uncapped claim of $2 million would contribute $100,000 to S_B. The contribution difference of $1.9 million is a structural truncation effect, separate from Laplace noise. The insurer might retain a separate tail model for decision support, but the release or training of that model must be reconciled with the privacy budget and provenance of the data used. A financial estimator can be privately released yet materially understate tail-risk capital if the analyst silently substitutes capped severity for the original target. The appropriate scientific output identifies both the protected query and the residual inference about the unrestricted outcome.
Composition, postprocessing, and adaptive use
A major advantage of differential privacy is its disciplined treatment of repeated use. If mechanisms are sequentially applied to the same protected individuals, a basic composition bound adds their epsilon and delta parameters. In more sophisticated schemes, privacy accounting may use tighter results under explicitly justified assumptions; the accountant and sampling design must be reproduced rather than guessed. Splitting one dataset into twenty dashboards does not create twenty independent privacy budgets. Distinct privacy units and disjoint participation groups may support different analyses, but their relationship must be proven from actual data lineages.
Postprocessing preserves an existing DP guarantee when the downstream operation uses only an already private output and independent public information. Recalibrating scores on original private individual outcomes, examining exact claims to select among repeated private models, or deciding which result to release based on sensitive data are additional operations requiring analysis. Interactive insurance analytics are particularly vulnerable to unbudgeted trial-and-error: the data scientist may repeatedly adjust categories, feature sets, thresholds, and reports until a rare subgroup becomes unusually informative. Query review and an enforceable privacy ledger are therefore part of technical correctness.
Learning algorithms and privacy accounting
The DP-SGD research of Abadi and colleagues (2016) develops private neural-network training through bounded per-example gradient contributions, added stochastic noise, and accounting for iterative access to the dataset. For person-level insurance learning, per-example gradient clipping alone may underprotect someone whose records appear multiple times; one must group contributions by the declared person, limit multiplicity, or derive a valid privacy bound for the actual sampling and participation scheme. Optimization convergence, choice of clipping norm, class imbalance, and batch structure determine practical utility.
A privacy accountant's output is only as sound as its assumptions. Whether sampling is independent or without replacement, whether all clients are observed, whether examples repeat across rounds, and whether a preliminary feature-selection step touched private data can change its validity. A report that advertises epsilon but omits delta, the privacy unit, population size, sampling protocol, and number of releases supplies insufficient information for scientific reproducibility. The correct output of an audit includes those details, the accountant implementation and its version, and the immutable execution configuration.
Local privacy and differentiated trust
In local differential privacy, individual contributions are randomized before the collection operator receives them. Erlingsson, Pihur, and Korolova (2014) examine a deployed approach in RAPPOR. Local privacy changes the trust boundary: the collector need not observe the original value for the specified report. Yet the privacy-utility tradeoff can be much harsher than in a trusted central mechanism, especially for sparse outcomes and high-dimensional combinations. A local randomized report also does not solve the problem of ensuring policy consent, coverage eligibility, or reliable claim attribution. The economic and supervisory decision may still require identifiable source records held separately under appropriate safeguards.
Theoretical private learnability research by Kasiviswanathan and colleagues (2011) relates learnability under local privacy to statistical-query learning under its formal assumptions. This demonstrates why the placement of the privacy mechanism is a fundamental statistical design choice. Local and central arrangements should not be compared by the epsilon number alone when their protected observations, data opportunities, and adversary models differ.
Federated and Distributed Modeling
Statistical object of collaboration
Federated learning allows organizations to compute parts of a shared learning procedure while retaining original records within separate administrative environments. In their federated optimization study, McMahan, Moore, Ramage, Hampson, and Agüera y Arcas (2017) introduce iterative model averaging and examine decentralized learning under unbalanced, heterogeneous data. The cross-organizational version is relevant to insurers, reinsurers, actuarial research consortia, and health-benefit administrators. Its statistical motivation may be the limited information available to any one organization about a rare risk category, exposure condition, or development period.
The economic motives for federation require scrutiny. Different insurers write different contracts and retain different risks. One may specialize in younger drivers, another in commercial fleets, and another in rural property. Their observed joint distribution of predictors and outcomes is altered by underwriting, marketing, policy limits, retention levels, claim-settlement practices, and legal jurisdictions. Combining models does not automatically recover a common target population. A federated project must state whether it seeks a pooled descriptive fit, a weighted industry estimator, a transportable predictive rule, or a model tailored to a specified participant's portfolio. These estimands need different sampling weights and validation.
Kairouz and colleagues (2021), in Advances and Open Problems in Federated Learning, document the wide range of statistical, systems, and privacy challenges in federated research. Their synthesis supports the characterization of federated learning as an architecture, not as a standalone formal protection guarantee. In insurance, independent data custodians may retain valuable governance control over claims, yet repeated gradient exchange can still disclose information. A claim of privacy preservation based solely on physical data locality would overstate what decentralization achieves.
Federated averaging and heterogeneous exposures
For K participating institutions, let n_k be the number of eligible training observations at institution k, and w_k a locally trained parameter vector. A common weighted aggregation is w = (sum_k n_k w_k)/(sum_k n_k). This expression is meaningful as an optimization step when each local procedure, weight, objective function, and reporting scale is correctly specified. It is not necessarily an estimator of the target insurer's prospective frequency or severity. A site with many short policies and a site with fewer long policies may require exposure weighting rather than raw row weighting. Repeated snapshots of the same claim can artificially increase n_k unless datasets are reconciled.
For a concrete illustration, suppose insurer A contributes 9,000 exposure units with a recorded count rate of 0.02 and insurer B contributes 1,000 units with a rate of 0.08. An exposure-weighted pooled descriptive frequency is (9,000 × 0.02 + 1,000 × 0.08)/10,000 = 0.026. The unweighted mean of institutional rates is 0.05, nearly twice as high. Neither quantity necessarily predicts the rate for a third insurer writing a different mix of business. The example involves no privacy noise: it isolates the statistical problem that remains even if distributed training is cryptographically secure and differentially private.
Client-level differential privacy can be defined over institutions; person-level privacy is different. A bound on the effect of dropping an insurer from training does not directly bound the effect of an individual claimant's record. Conversely, person-level protection needs a trustworthy process for per-person contributions across clients. A policyholder may switch carriers, hold overlapping coverages, or appear in several subsidiaries. Global deduplication could itself require privacy-preserving linkage and legal authorization. The participating institutions should document the actual protected unit and identify whether cross-client multiplicity is bounded or accepted as a limitation.
Secure aggregation: function and limits
Bonawitz and collaborators (2017), in Practical Secure Aggregation, develop a protocol enabling a server to obtain an aggregate of client updates without observing each individual plaintext update under stated protocol assumptions. Such cryptographic aggregation can reduce what an infrastructure operator learns about participants. Its benefits depend on honest-but-curious or malicious behavior assumptions, dropout handling, key management, participant numbers, and the absence of unauthorized alternative channels. It is an important confidentiality instrument, but the aggregate itself may still reveal information about a small contributing group or permit comparisons across rounds.
Consider four insurers whose clipped scalar updates are x1, x2, x3, and x4. Secure aggregation reveals x1+x2+x3+x4 rather than the four components. If one institution's contribution is known through another channel, the remaining sum becomes more informative; if successive rounds involve changing subsets, differencing may expose components. A protocol that protects transmitted updates is not a proof that the final model is differentially private with respect to individual persons. Private noise, contribution bounding, output access limits, and composition analysis may remain necessary. The threat model must therefore identify both the computation that is hidden and the output that is revealed.
Distributed differential privacy introduces additional questions. If participants jointly produce one noisy sum, the aggregate noise must have a proven distribution appropriate for the global sensitivity and target privacy guarantee. Arbitrarily adding the same amount of local noise at every site does not in general implement a specified central mechanism. A malicious participant may suppress noise, change the roster, or adapt its contribution to learn about others. Protocol designs must explicitly address collusion thresholds, dropout recovery, and trusted randomness. The cryptographic security theorem and the statistical privacy theorem may be proved under different assumptions; the system-level argument needs their intersection.
Model update leakage and inference services
The gradient leakage demonstrated by Zhu and colleagues (2019) makes the communication surface of federated training a scientific object. A few organizations exchanging gradients in small batches may create a different disclosure exposure from millions of independent clients supplying clipped updates. An unusual claimant can dominate a rare-class gradient. Even if the output is protected by secure aggregation, participants may inspect the shared model repeatedly and attempt membership or attribute inference. Whether differential privacy protects each individual depends on what is randomized, when and where it is randomized, the number of rounds, and the chosen adjacency unit.
Papernot and colleagues (2018), in Scalable Private Learning with PATE, examine private transfer of knowledge from models trained on disjoint private partitions to a student model through noisy teacher votes. This illustrates another separation between confidential training data and a publishable model. Its mechanisms are not automatically suitable for every continuous actuarial outcome: claims severity, survival time, reserve distributions, and continuous pricing functions raise tasks different from the classification settings reported in that study. Private prediction services may require additional work on output formats, calibration, uncertainty, and the total number of adaptive queries.
Institutional architecture and legal boundaries
A federation of insurers may cross national borders, involve personal or health information, use third-party computing vendors, and process records collected under different contracts. Keeping source records in their original jurisdiction can reduce some movement of data but does not by itself settle whether model updates, gradients, identifiers, or other outputs are personal data under applicable law. Legal assessment requires the actual information content and risk of identification, along with controller and processor roles, sharing agreements, permitted purposes, and cross-border transfer requirements. Contractual claims about anonymity should follow a technical disclosure assessment rather than precede it.
A credible cross-insurer collaboration should retain written responsibility for data definitions, selection criteria, identifiers, contribution caps, local validation, cryptographic protocols, threshold parameters, privacy accounting, incident response, and withdrawal rules. Model governance also needs participant-specific checks. If a shared predictive rule performs adequately in the largest participant's portfolio but fails for the smaller carrier's claims, the collaboration has created an operational risk even where its aggregate privacy guarantee is mathematically correct. Federated statistical learning is successful only when its target population, privacy unit, and deployment context are jointly coherent.
Privacy-Utility Tradeoffs
Utility is a decision-relative quantity
Privacy mechanisms introduce randomness, truncation, approximation, or restrictions on training. Their utility should be measured against the purpose of the insurance model rather than against a single generic accuracy metric. For claim frequency, an error of 0.005 might be negligible in a common class and consequential in a rare high-risk cohort. For reserve estimation, distortion in the tails or the development pattern may be more important than average squared error. For underwriting classification, calibration and stability within legally permissible risk classes may matter more than overall ranking accuracy. For mortality, a small error at old ages can accumulate into material present-value differences for long-duration annuities.
A privacy-utility curve is meaningful only after identifying the estimand, sample unit, mechanism, privacy definition, model family, exposure basis, and evaluation population. Varying epsilon without holding those conditions fixed conflates methodological choices. The statistical target may change when claims are clipped or when information about rare diagnoses is suppressed. Comparing a capped private estimator with an uncapped nonprivate benchmark can show overall decision loss, but reporting an inference error solely relative to the capped target conceals the difference between the final estimate and the contractual financial quantity. Both comparisons are useful and should be retained.
Sampling variation and mechanism variation
Let Y be an observed claim count and Z independent mean-zero private noise. With a fixed exposure E, the released estimate is R=(Y+Z)/E. If Y follows a Poisson law with mean lambda E and Z is Laplace with scale b, then E[R]=lambda and Var(R)=lambda/E + 2b^2/E^2, conditional on the assumptions and absent postprocessing. The first term reflects stochastic claim-count variability; the second reflects private randomization. For a realized count y, one may estimate the first term by y/E^2 in a simple plug-in analysis, while recognizing estimation uncertainty and model misspecification. This decomposition clarifies why privacy noise may be secondary at high exposure and primary for narrow groups.
For a separate illustrative Poisson count model, suppose lambda=0.1, E=10,000, b=2; this process assumption differs from the preceding binary-indicator dataset. Then process-rate variance is 0.1/10,000 = 0.00001, and privacy-rate variance is 8/100,000,000 = 0.00000008. The resulting model-based standard deviation is sqrt(0.00001008), approximately 0.003175, while the privacy-only standard deviation remains 0.000283. For a rare subgroup with lambda=0.02 and E=50, the process-rate variance is 0.0004, and the privacy-rate variance is 0.0032, yielding total standard deviation sqrt(0.0036)=0.06. The private component then drives most of the total variation. This stylized result makes no statement about the distribution of real claims, dependencies among insureds, or the quality of the rate estimator under product changes.
Bias and postprocessing
A released noisy count may be constrained to the set of nonnegative integers or a proportion to [0,1]. Postprocessing does not increase the information revealed from the already private output, provided it uses no additional private data. However, rounding, clipping, or thresholding changes expected values and error distribution, especially near zero. A claim count with true value zero and symmetric noise becomes positively biased if negative outputs are replaced with zero. The same operation can compress distinctions among rare but different risk classes. An audit should therefore report both the unmodified mechanism and any downstream transformation.
In a model with exposure E and count N, replacing negative private N with zero produces a floor at zero. The frequency estimator becomes max(0, N+Z)/E. Its expectation exceeds N/E for any continuous symmetric noise with positive probability below -N, and the size of the bias is most material when N is small relative to the noise scale. For pricing, additional constraints may be imposed by filed rating structures, product design, consumer protection, or professional judgment. Such adjustments should be traceable and analyzed for distributional consequences instead of being described as if they followed directly from the privacy theorem.
Tail risk, rare events, and credibility
Insurance inference repeatedly confronts small samples with large consequences. A very large claim may be the most informative observation for catastrophe coverage or excess-of-loss pricing, precisely where contribution caps suppress its magnitude. Long reporting delays can make paid-loss triangles sparse by development period, and sensitive claims histories may contain rare facts that make even summary data linkable. Private learning in these settings requires explicit bias analysis, not a generic statement that adding noise preserves approximate accuracy. The scientific decision may be to release a coarser estimand, use public covariates, pool legitimately comparable groups, or withhold a statistic until sufficient exposure accumulates.
Experience rating and credibility address sparse information through structured borrowing of strength, while privacy mechanisms restrict the influence of any one person's contribution. These processes may coexist, but their objectives differ. A stronger prior or hierarchical shrinkage can stabilize an estimate under noise, yet it introduces its own assumptions about comparability and heterogeneity. Shrinkage learned from the same confidential experience may itself incur privacy cost. If one subgroup is particularly rare, an apparently stable pooled estimate can be driven by the much larger group and produce unacceptable conditional error. The relevant question is whether the privacy-preserving estimator serves the predefined insurance decision for each intended population.
Calibration, fairness, and distribution shift
Overall predictive discrimination is insufficient for sensitive insurance uses. Calibration asks whether the predicted quantity corresponds to observed incidence or cost in defined groups. Protected or vulnerable subgroups may have different exposures, claim propensities, reporting delays, and observation quality. Differentially private training may affect those groups unevenly when the contribution bound or added noise interacts with sample size. Measuring disparity can itself require protected group labels and sensitive outcomes; such audit statistics may need their own privacy accounting. The audit therefore must document whether privacy parameters protect validation cohorts as well as the training records.
A model trained before a coverage reform or a major disease episode may be private yet invalid after the distribution changes. Reuse of model parameters in a new period does not repair changed underwriting, exposure, medical practice, or judicial interpretation. A retraining schedule introduces additional repeated access to individuals' histories and may increase cumulative privacy cost if releases share a protected population. The required evidence includes time-split validation, performance by insurer and contract class, stress cases, drift detection, and the privacy cost of model replacement. Privacy and model lifecycle management are coupled through repeated observation, even though their formal test criteria remain distinct.
Comparing architectures without spurious rankings
A trusted central DP estimator, a locally private data-collection system, an encrypted central learner, and a federated learner with secure aggregation can all be legitimate designs. No unconditional ranking follows from their labels. Central DP can provide high statistical efficiency for a specified release, while requiring a trusted operator with access to originals. Local DP protects against collection-side disclosure for the specified reports, often at higher statistical cost. Encryption protects transit and storage surfaces but does not itself limit inference from outputs. Federated learning reduces the movement of raw records while introducing update and coordination surfaces. Secure aggregation reduces intermediate visibility while leaving the aggregate and model accessible under its protocol. A combined system has to establish that the assumptions behind every layer are simultaneously satisfied.
This comparison identifies the appropriate scientific standard for a privacy-utility claim. It must describe the threatened information, protected unit, exact mechanism, formal parameters if any, data distribution, decision loss, operational dependencies, and unprotected surfaces. The finding that one technique was useful in another domain, or that a vendor advertises anonymization, provides no substitute for this evidence. Quantitative comparisons should preserve a common target, population, evaluation period, and privacy requirement; otherwise, apparently superior results may be produced by a changed problem definition.
Applicability and Audit Methods
Mapping to insurance functions
For ordinary frequency analysis, a bounded count over a publicly known exposure denominator can be a transparent starting point, with mechanism variance and count-process variance reported separately. For severity, the tail-cap problem requires a contractual explanation and a documented approximation for the portion of loss excluded from sensitivity-bounded totals. For claims reserving, private development factors and small triangle cells may be sensitive to the amount of noise and to correlated outputs across periods. For life and health insurance, strong person-level protection can be important for medical histories and causes of death; individualized pricing and consumer-protection duties make validation of errors and treatment of sensitive attributes especially significant. For fraud analysis, sharing patterns across carriers can aid detection but creates considerable sensitivity around investigations that never resulted in confirmed fraud.
In catastrophe analytics, large aggregate exposure clusters and correlated claims can make risk units difficult to define. An individual insured's marginal contribution may be economically substantial, while the institutional need may concern confidential insured locations rather than a person's personal data. Differential privacy at person level may be irrelevant to a commercial-property portfolio where the protected object is a factory's location or a firm's trading exposure. The adjacent dataset should be chosen from the actual confidentiality obligation, and group or enterprise privacy may require a distinct analysis. Privacy engineering for insurance therefore includes both personal-data and commercially confidential-data research, with separate threat models.
International supervisory distinction
The EU's GDPR regulation is a binding legislative instrument within its applicable scope. Technical minimization and differential privacy can assist compliance assessments, but lawful processing still turns on legal bases, purpose, special-category rules, individual rights, controller responsibilities, security, and transfer conditions. A supposedly anonymous output should be assessed in context, including reasonably available means of identification; one should not infer a legal classification merely because a mechanism has a named mathematical property. The technical guarantee and the legal finding use different criteria and may need different evidence.
The European Insurance and Occupational Pensions Authority's 2025 AI governance Opinion is a supervisory opinion addressed to national authorities. EIOPA explains that it clarifies existing insurance-sector governance and risk-management principles rather than creating new requirements. Its official explanatory release distinguishes insurance uses categorized as high risk or prohibited under the EU AI Act from the Opinion's own scope, and notes that AI used for risk assessment and pricing in life and health insurance is treated as high risk under the Act. The dates of application and the precise legal classification of any deployed model should be determined against the applicable instrument at the relevant time. No general conclusion that every actuarial statistical model is high-risk follows from this institutional summary.
In the United States, the National Association of Insurance Commissioners' 2023 AI Model Bulletin was approved as model guidance for state regulators. NAIC expressly describes it as guidance and not a model law or regulation. State-level adoption and the exact operative bulletin differ by jurisdiction and date. Its expectations concerning accuracy, unfair discrimination, governance, documentation, and third-party models provide an important framework for insurer use of AI, but a national, directly binding rule should not be inferred from the NAIC vote alone. The United States ASOP No. 56 concerns professional model work within its applicability, not legal permission to process any personal records.
International supervisory bodies and insurers have different institutional roles. A standard authored by a US actuarial professional body should not be described as EU law; an EU supervisory opinion should not be imported as an obligation for every market; a technical NIST publication is guidance rather than legislation. The scientific objective of international comparison is to identify how privacy design can coexist with applicable duties, while retaining local legal specificity and avoiding unsupported harmonization.
An independently reproducible validation protocol
A complete validation process starts with a versioned research question specifying theta, exposure measure, outcome definition, forecast horizon, contract scope, and intended use. The data inventory should identify persons, duplicate claims, sensitive categories, observation windows, access controls, and participation across internal and external systems. A threat register then records what each adversary can see and what private fact that adversary seeks to infer. For each mechanism, the analyst documents the neighboring dataset relation, sensitivity calculation, contribution constraints, randomization procedure, privacy parameters, composition rules, and the details of any secure computation. The method statement must say which protections are formal proofs and which are engineering or organizational mitigations.
Evaluation follows the defined statistical target. A nonprivate reference estimator is built where authorized, and a private estimator is compared on synthetic or appropriately governed validation data. The assessment should distinguish process variation, parameter uncertainty, privacy randomization, model misspecification, selection bias, and release postprocessing. It should examine the full portfolio and prespecified subgroups, including rare strata. Both point and distributional error should be considered, with calibration, tail metrics, and stability by time and participant. A single test set evaluated many times may become a privacy leakage surface or an unreliable selector if model choice repeatedly adapts to confidential outcomes; the test protocol and its accounting should therefore be specified in advance.
A useful computational audit includes tests for the declared contribution cap, independent random-number generation, sensitivity scale, noise variance, composition arithmetic, reproducible seeds in nonproduction synthetic tests, bounded actual disclosure of the generated outputs, and adherence to trained-model release limits. Public reproducibility should rely on synthetic examples and published configurations rather than reproducing private records. An independent reviewer should be able to recompute the illustrative count and loss calculations above without access to any insured person. A stronger validation claim would require controlled real-world evaluation and documented authority to process the relevant records.
Interpretability and contestability
A model may be privately trained but insufficiently documented for underwriting, rate filings, individual claim decisions, or consumer recourse. A privacy guarantee about training participation does not establish causal validity, explainability of individual scores, or fairness of protected-group outcomes. The institution should identify which explanation is required for which actor: a statistical explanation for researchers, a model-governance record for supervisors, a contractual rationale for insureds, or a decision-specific communication for affected applicants. These can be served through different information channels, provided none is misleading about the model's operation or the data actually used.
Overly detailed explanations could themselves disclose training information if they expose nearest neighbors, raw examples, or sensitive gradients. Conversely, privacy claims should not be invoked as a blanket reason to omit the information required by applicable governance and consumer protections. Information-access decisions need their own proportionate assessment. A privacy-preserving model can remain auditable through versioned documentation, approved aggregate metrics, authorized secure examination, and procedures for handling challenges. The scope of these controls depends on the decision and the jurisdiction.
Failure modes and correction pathways
Scientific failure modes include an incorrect protected-unit definition, unbounded individual contributions, optimistic assumptions about sampling amplification, omission of hyperparameter selection from privacy accounting, mismatched cryptographic and adversary assumptions, repeated uncounted releases, gradient or model-output leakage, implausible denominators, misuse of capped losses as uncapped predictions, and unjustified transfer between insurer portfolios. Operational failures include unexpected data joins, insufficient audit logs, nonreproducible software versions, and authorization granted to a partner beyond the original purpose. A release should be paused when a material privacy proof assumption is invalid or when decision accuracy is insufficient for its stated actuarial use. The reason and evidence should be documented, and the system should be recalibrated or redesigned before renewed assessment.
The correction process is not exhausted by changing epsilon. The proper remedy may involve changing the unit of protection, limiting query frequency, aggregating categories, redesigning the statistical estimand, reducing features, adding secure aggregation, rotating credentials, altering consortium participation, or withholding a specific subgroup statistic. Every change can affect financial and legal interpretation. A change log must therefore relate each revised protection configuration to the corresponding research and production model version. The final privacy statement describes the exact released version and its applicable guarantees, not an ideal algorithm detached from implementation.
Scholarly synthesis and reproducibility limits
The collected literature supports a hierarchy of defensible statements. Differential privacy provides a formal conditional output guarantee; secure aggregation provides a protocol-dependent confidentiality property of intermediate contributions; federated learning provides a distributed training architecture; actuarial standards provide professional expectations for data and model quality; statutes and supervisory instruments impose or interpret duties within their own jurisdictions. The substantive scientific connection is that the utility of a privacy-preserving insurance result must be assessed relative to a defined actuarial estimand and the insured-person contribution boundary, rather than merely against a generic benchmark score. Once that boundary and estimand are explicit, the analyst can determine which outputs require protection, which privacy costs compose, and whether the resulting noise materially changes the insurance decision.
This is an original analytic synthesis of established research and institutional sources. It makes no claim to a new formal differential privacy theorem, no newly established actuarial doctrine, and no empirical performance result from real insurer data. The illustrative calculations establish algebraic consequences under declared assumptions. They do not establish adequate privacy parameters for any real portfolio. Future empirical investigation could compare person-level DP, local DP, and secure federated training on authorized insurer data with pre-registered threat models, matched estimands, subgroup calibration, and auditably recorded releases. Such work would be a separate study whose results could modify the practical conclusions offered here.
Conclusion
Privacy-preserving statistical learning for insurance is the scientific study of how sensitive policyholder and claim experience can inform estimators and predictive models while the contribution of protected information to released results is demonstrably controlled. Its core questions are statistical and institutional: what is being estimated, which person or entity is protected, which outputs an adversary may observe, which mechanism bounds disclosure, how repeated access composes, and whether the remaining estimate supports the contractually relevant decision. A privacy statement that answers only how raw files are stored does not answer these questions.
Formal differential privacy provides a strong language for bounding marginal participation effects under specified assumptions. The contribution bound, neighboring-dataset relation, sensitivity, epsilon, delta, accountant, and release history are inseparable parts of that language. The hypothetical examples show that a single count-noise distribution can be minor for 10,000 exposure units and dominant for 50, while claims-loss clipping can make an estimator privately releasable by replacing the unrestricted tail-dependent target with a bounded one. These are actuarially decisive distinctions.
Distributed learning can preserve the institutional locality of original records, and secure aggregation can reduce exposure of individual model updates, but neither property by itself guarantees that shared gradients or a final model cannot reveal membership or sensitive facts. A correct system may combine trustworthy data governance, bounded private computation, cryptographic protocols, documented model selection, and limited release interfaces. Its guarantees must match the full information flows and the plausible adversary.
International regulatory and professional sources reinforce the need for differentiated controls. EU data-protection legislation, EIOPA supervisory interpretation, the US NAIC model framework, and US actuarial standards have distinct legal and professional functions. Privacy mechanisms are evaluated alongside data quality, fairness, calibration, consumer rights, and model governance without collapsing these obligations into a single metric. The final scientific conclusion is that reliable privacy-preserving insurance learning requires the protection unit, statistical estimand, release design, and decision validation to be specified and audited together. That conclusion follows from the analyzed literature and mathematical derivations, and is offered as an analytical synthesis rather than a new named theory.
References
Abadi, M., Chu, A., Goodfellow, I., McMahan, H. B., Mironov, I., Talwar, K., and Zhang, L. (2016). Deep Learning with Differential Privacy. Proceedings of the 23rd ACM SIGSAC Conference on Computer and Communications Security, 308–318. Original publisher research record.
Actuarial Standards Board. (2016). Actuarial Standard of Practice No. 23: Data Quality (December 2016 revision). United States actuarial professional standard. Official standard.
Actuarial Standards Board. (2019). Actuarial Standard of Practice No. 56: Modeling (December 2019 revision; effective October 1, 2020). United States actuarial professional standard. Official standard.
Bonawitz, K. A., Ivanov, V., Kreuter, B., Marcedone, A., McMahan, H. B., Patel, S., Ramage, D., Segal, A., and Seth, K. (2017). Practical Secure Aggregation for Privacy-Preserving Machine Learning. Proceedings of the 2017 ACM SIGSAC Conference on Computer and Communications Security, 1175–1191. DOI: 10.1145/3133956.3133982. Research record.
Dwork, C. (2006). Differential Privacy. In M. Bugliesi, B. Preneel, V. Sassone, and I. Wegener (eds.), Automata, Languages and Programming, ICALP 2006, Part II. Springer. Author's institutional research record.
Dwork, C., McSherry, F., Nissim, K., and Smith, A. (2006). Calibrating Noise to Sensitivity in Private Data Analysis. In S. Halevi and T. Rabin (eds.), Theory of Cryptography, TCC 2006, Lecture Notes in Computer Science 3876. Springer. DOI: 10.1007/11681878_14. Publisher DOI record.
Dwork, C., and Roth, A. (2014). The Algorithmic Foundations of Differential Privacy. Foundations and Trends in Theoretical Computer Science, 9(3–4), 211–487. DOI: 10.1561/0400000042. Publisher DOI record.
Erlingsson, Ú., Pihur, V., and Korolova, A. (2014). RAPPOR: Randomized Aggregatable Privacy-Preserving Ordinal Response. Proceedings of the 21st ACM SIGSAC Conference on Computer and Communications Security, 1054–1067. Original research record.
European Insurance and Occupational Pensions Authority. (2025a). Opinion on Artificial Intelligence Governance and Risk Management. EIOPA-BoS-25-360, August 6, 2025. Supervisory opinion to national competent authorities. Official opinion record.
European Insurance and Occupational Pensions Authority. (2025b). EIOPA Publishes Opinion on AI Governance and Risk Management. News release, August 6, 2025. Official explanatory release.
European Parliament and Council. (2016). Regulation (EU) 2016/679 of April 27, 2016 (General Data Protection Regulation). Official Journal of the European Union, L 119, 1–88. Official EUR-Lex legislative text.
Kairouz, P., McMahan, H. B., et al. (2021). Advances and Open Problems in Federated Learning. Foundations and Trends in Machine Learning, 14(1–2), 1–210. DOI: 10.1561/2200000083. Publisher DOI record.
Kasiviswanathan, S. P., Lee, H. K., Nissim, K., Raskhodnikova, S., and Smith, A. (2011). What Can We Learn Privately? SIAM Journal on Computing, 40(3), 793–826. DOI: 10.1137/090756090. Publisher DOI record.
McMahan, H. B., Moore, E., Ramage, D., Hampson, S., and Agüera y Arcas, B. (2017). Communication-Efficient Learning of Deep Networks from Decentralized Data. Proceedings of the 20th International Conference on Artificial Intelligence and Statistics, Proceedings of Machine Learning Research, 54, 1273–1282. Original conference publication.
National Association of Insurance Commissioners. (2023). NAIC Members Approve Model Bulletin on Use of AI by Insurers. December 4, 2023. US insurance regulators' model guidance adoption announcement. Official NAIC announcement.
National Institute of Standards and Technology. (2020). NIST Privacy Framework: A Tool for Improving Privacy Through Enterprise Risk Management, Version 1.0. Voluntary framework; official page also documents later draft development. NIST framework portal.
Near, J. P., Darais, D., Lefkovitz, N., and Howarth, G. S. (2025). Guidelines for Evaluating Differential Privacy Guarantees. NIST Special Publication 800-226, March 2025. DOI: 10.6028/NIST.SP.800-226. NIST final publication.
Papernot, N., Song, S., Mironov, I., Raghunathan, A., Talwar, K., and Erlingsson, Ú. (2018). Scalable Private Learning with PATE. International Conference on Learning Representations. Original research record.
Shokri, R., Stronati, M., Song, C., and Shmatikov, V. (2017). Membership Inference Attacks Against Machine Learning Models. 2017 IEEE Symposium on Security and Privacy, 3–18. DOI: 10.1109/SP.2017.41. Author's original paper record.
Zhu, L., Liu, Z., and Han, S. (2019). Deep Leakage from Gradients. Advances in Neural Information Processing Systems, 32. Original conference paper.