{
  "id": "differential-privacy",
  "title": "Differential Privacy",
  "lang": "en",
  "source_lang": "en",
  "summary": "Differential privacy is a mathematical framework for limiting how much an analysis can reveal about any one individual’s data.",
  "category": "computer-science",
  "subcategory": "networks-security",
  "type": "concept",
  "aliases": [
    "DP"
  ],
  "generated_at": "2026-11-03",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/differential-privacy/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/differential-privacy/",
      "json": "https://aiwiki.page/zh/differential-privacy.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Differential privacy is a mathematical framework for analyzing data while limiting what the analysis reveals about any particular individual. A differentially private randomized algorithm produces outputs whose probability distributions change only within specified bounds when one person’s data is added, removed, or altered. This makes it possible to publish useful population statistics or train predictive models without making the result overly dependent on a single participant.\n\nThe framework occupies a central place in data privacy because its guarantee is expressed independently of a particular attacker’s knowledge or a particular method of identifying people. Removing names from records does not necessarily prevent identification, and releasing many apparently harmless statistics can reveal underlying records. Differential privacy instead constrains the information conveyed by the release itself. Its guarantees apply to a specified mechanism, a specified unit of protection, and explicit numerical privacy parameters—not automatically to an entire organization or data-processing system.\n\nSince its formulation in 2006, differential privacy has developed into a substantial area of computer science and statistics. Its theory includes mechanisms for adding calibrated noise, selecting useful outputs, combining repeated analyses, and allocating privacy loss. Applications include official statistics, software telemetry, database query systems, and machine learning. The principal practical challenge is to preserve enough statistical accuracy while faithfully implementing the assumptions under which the privacy guarantee holds."
    },
    {
      "heading": "Origins and the failure of anonymization",
      "text": "Traditional statistical disclosure control includes suppressing small counts, rounding values, removing direct identifiers, and replacing detailed attributes with broader categories. These methods can reduce disclosure risk, but their effectiveness depends strongly on what other information is available. A record without a name may still be identifiable through a combination of location, age, occupation, or other attributes. Linking such a record to an external source can reveal information that the original publisher intended to conceal. Consequently, anonymity is not simply a property of whether a table contains an explicit identity column.\n\nA deeper problem arises when releases contain many overlapping statistics. Each statistic constrains the possible underlying dataset. Enough accurate constraints can allow an attacker to reconstruct much of that dataset, even if no individual record was directly published. In 2003, Irit Dinur and Kobbi Nissim established an influential reconstruction result showing that sufficiently accurate answers to sufficiently many subset-counting queries can expose most of a binary database. The importance of the result was not that every aggregate release is unsafe, but that small amounts of noise do not necessarily remain protective when the release is sufficiently extensive.\n\nDifferential privacy emerged from research on this tension between usefulness and disclosure. Cynthia Dwork, Frank McSherry, Kobbi Nissim, and Adam Smith introduced the framework under its now-standard name in their 2006 paper “Calibrating Noise to Sensitivity in Private Data Analysis.” Their approach connected the amount of random noise required to the maximum effect that one individual could have on the requested answer. Earlier work on an indistinguishability-based definition supplied part of the conceptual foundation. Subsequent research developed the framework far beyond numerical queries, including private selection, optimization, learning, and continual data analysis.\n\nThe shift was from trying to classify a released dataset as anonymous to specifying how an analysis behaves under neighboring inputs. This distinction also separated differential privacy from encryption. Encryption can protect stored or transmitted records from unauthorized access, but an authorized recipient may still infer sensitive information from a decrypted statistical result. Differential privacy addresses the informational consequences of releasing that result. The two techniques can therefore be complementary rather than competing solutions.\n\nThe historical motivation also explains why differential privacy does not promise that no sensitive fact can ever be inferred. Population-level regularities can support accurate predictions about individuals, even when those individuals did not contribute data. The framework instead restricts the additional disclosure attributable to one protected contribution. Its strongest claims concern this incremental influence, not the elimination of every inference that statistical knowledge makes possible."
    },
    {
      "heading": "Mathematical definition and the unit of protection",
      "text": "Let \\(D\\) and \\(D'\\) be neighboring datasets, written \\(D\\sim D'\\), and let \\(M\\) be a randomized mechanism. The mechanism is \\((\\varepsilon,\\delta)\\)-differentially private if, for every neighboring pair and every measurable set of outputs \\(S\\),\n\n\\[\\Pr[M(D)\\in S] \\leq e^\\varepsilon\\Pr[M(D')\\in S]+\\delta.\\]\n\nThe probabilities are over the mechanism’s internal randomness. The requirement applies to every allowed dataset, not merely to datasets that appear likely under a statistical model. Because the neighboring relation is normally symmetric, the inequality also applies with \\(D\\) and \\(D'\\) exchanged. Thus neither dataset can make an output event arbitrarily more likely, except within the allowance represented by \\(\\delta\\).\n\nThe parameter \\(\\varepsilon\\), pronounced “epsilon,” controls the multiplicative difference between output probabilities. Smaller values impose a stronger constraint. When \\(\\delta=0\\), the guarantee is called pure differential privacy, commonly abbreviated \\(\\varepsilon\\)-differential privacy. When \\(\\delta>0\\), it is called approximate differential privacy. The additive term allows limited departures from the pure multiplicative bound. It is not generally correct to interpret \\(\\delta\\) as the probability that the entire dataset will be disclosed, or \\(\\varepsilon\\) as a direct percentage probability of identifying someone.\n\nThe definition depends critically on what “neighboring” means. Under add-or-remove adjacency, one dataset contains one additional protected contribution. Under substitution adjacency, one contribution is replaced by another. These conventions yield different sensitivities and may require different parameter calibrations. A contribution might be a database row, all records associated with one person, a household, or an organization. Event-level protection for a single location measurement is substantially different from user-level protection for an entire location history.\n\nFor example, if each person contributes one survey response, row-level and person-level protection may coincide. If one person contributes thousands of purchase records, protecting each row separately does not automatically protect the person’s complete purchase history at the same parameters. Contribution limits, identity resolution, and grouping rules therefore form part of the privacy specification. A statement that a system is “differentially private” is incomplete unless the protected unit and neighboring relation are understood.\n\nA simple group guarantee illustrates this dependence. If a mechanism is \\(\\varepsilon\\)-differentially private for one record, changing up to \\(k\\) records gives a bound of \\(k\\varepsilon\\) under the corresponding chain of neighboring datasets. Approximate differential privacy has an additional accumulated \\(\\delta\\) term. This group privacy property is useful, but it also shows why weakly bounded participation can undermine a nominally strong record-level guarantee. The protected unit is a mathematical and operational choice, not a detail that can be inferred from the algorithm’s name."
    },
    {
      "heading": "Sensitivity and the principal mechanisms",
      "text": "The most familiar mechanisms add noise calibrated to a query’s sensitivity. For a numerical function \\(f\\), its global \\(L_1\\) sensitivity is\n\n\\[\\Delta_1 f = \\sup_{D\\sim D'} \\lVert f(D)-f(D')\\rVert_1.\\]\n\nThis measures the largest possible change in the answer between neighboring datasets. A count of people satisfying a condition has sensitivity \\(1\\) under add-or-remove adjacency when each person contributes at most once. For a histogram in which each person belongs to one bin, adding a person changes one bin by \\(1\\); replacing a person can change two bins.\n\nThe Laplace mechanism releases\n\n\\[M(D)=f(D)+Z, \\qquad Z\\sim\\operatorname{Laplace}(0,\\Delta_1 f/\\varepsilon)\\]\n\nfor a scalar query, with independent coordinate noise providing the corresponding vector construction. The Laplace distribution has density proportional to \\(\\exp(-|z|/b)\\), where \\(b\\) is its scale. Its shape ensures that shifting the true answer by at most the sensitivity changes output likelihoods by no more than the permitted multiplicative factor. For a sensitivity-\\(1\\) count, \\(\\varepsilon=1\\) gives noise scale \\(1\\), while \\(\\varepsilon=0.1\\) gives scale \\(10\\). The expected absolute error equals that scale.\n\nThe Gaussian mechanism adds normally distributed noise and generally provides approximate rather than pure differential privacy. Its calibration uses \\(L_2\\) sensitivity. A classical sufficient condition, for \\(0<\\varepsilon<1\\), is\n\n\\[\\sigma \\geq \\frac{\\Delta_2 f\\sqrt{2\\ln(1.25/\\delta)}}{\\varepsilon}.\\]\n\nThis is a convenient bound rather than the tightest possible calibration. More precise methods evaluate the privacy properties of Gaussian noise directly. Gaussian mechanisms are especially important in high-dimensional statistics and iterative learning, where their composition can be analyzed efficiently.\n\nSensitivity often has to be enforced rather than merely calculated. A sum of unbounded incomes has unbounded sensitivity. Clipping each contribution to a fixed interval creates a finite bound, at the cost of changing the quantity being estimated. A mean requires attention to both its numerator and denominator: a bounded sum divided by a public fixed count differs from a ratio involving a privately estimated count. Likewise, a database join may multiply a person’s contribution across many output rows. These structural effects can dominate the noise scale.\n\nNot every private release is best expressed as a noisy number. The exponential mechanism, introduced by McSherry and Kunal Talwar in 2007, selects an output \\(r\\) according to a quality score \\(u(D,r)\\):\n\n\\[\\Pr[M(D)=r] \\propto \\exp\\!\\left( \\frac{\\varepsilon u(D,r)}{2\\Delta u} \\right),\\]\n\nunder the usual discrete-output formulation. It favors high-quality choices while limiting the influence of one participant on their selection. Applications include choosing a model, a threshold, or a category. More specialized mechanisms use smooth sensitivity, threshold tests, or carefully correlated noise when generic independent noise would waste accuracy."
    },
    {
      "heading": "Composition, post-processing, and privacy accounting",
      "text": "Differential privacy is designed to support reasoning about whole sequences of analyses. Its post-processing theorem states that transforming a private output without further access to the protected data does not weaken its privacy guarantee. A noisy table may be rounded, visualized, projected onto consistency constraints, or used to train another model without an additional privacy charge arising solely from that processing. The transformation may use public information or independent randomness. If it accesses the confidential dataset again, however, that access must receive its own analysis.\n\nComposition addresses repeated access. If mechanisms \\(M_1,\\ldots,M_k\\) provide \\((\\varepsilon_i,\\delta_i)\\)-differential privacy, their joint release satisfies the basic bound\n\n\\[\\varepsilon_{\\mathrm{total}} = \\sum_{i=1}^{k}\\varepsilon_i, \\qquad \\delta_{\\mathrm{total}} = \\sum_{i=1}^{k}\\delta_i.\\]\n\nThis remains valid for adaptive analyses when each mechanism satisfies the required conditional guarantee given the previous transcript. An analyst may choose the next query after seeing earlier answers, but the privacy cost does not disappear because the queries were exploratory. For pure differential privacy, repeated identical queries with fresh independent noise also accumulate privacy loss; averaging their answers progressively reduces the noise.\n\nThe term privacy budget describes the resource tracked by such composition rules. It is not a quantity of privacy that can be measured directly in a person’s life. Rather, it records a bound on the disclosure permitted by a particular collection of mechanisms. A system might allocate part of a budget to selecting categories, part to estimating counts, and part to validating a model. If the same already-published private output is distributed again, that redistribution does not itself create a new independent observation of the underlying data.\n\nBasic addition is often conservative. Advanced composition can yield substantially smaller bounds for long sequences of mechanisms. Rényi differential privacy, introduced by Ilya Mironov in 2017, expresses privacy through Rényi divergences between neighboring output distributions. It is particularly useful because costs at a fixed divergence order add under composition. The resulting bound can then be converted into an \\((\\varepsilon,\\delta)\\) statement. Concentrated differential privacy and privacy-loss-distribution methods provide other representations of accumulated privacy loss.\n\nSubsampling can also amplify privacy under appropriate assumptions. If an individual is randomly included with probability \\(q\\), the overall mechanism may reveal less about that individual than the mechanism applied to a dataset where inclusion is certain. The exact amplification formula depends on the sampling procedure, neighboring relation, and privacy definition. Poisson sampling, fixed-size sampling, and sampling with replacement cannot be interchanged casually in an accountant. Similarly, parallel composition can avoid adding costs for disjoint groups, but only when those groups are disjoint with respect to the actual protected unit."
    },
    {
      "heading": "Central, local, and distributed trust models",
      "text": "The central model assumes that a trusted curator can access the confidential data and releases only differentially private results. A statistical agency or database service may operate in this way. The guarantee protects against what recipients can learn from the released outputs; it does not prevent the curator from seeing the original records. Access control, secure storage, auditing, and organizational procedures remain necessary to protect that internal access. Central mechanisms often achieve high accuracy because they add noise after aggregating many contributions.\n\nIn local differential privacy, each participant randomizes a contribution before the collector receives it. A local randomizer \\(R\\) satisfies a likelihood-ratio bound across possible input values, so the collector does not need to receive an unperturbed value. A foundational example is randomized response, introduced by Stanley Warner in 1965, decades before differential privacy was named. A respondent may answer a sensitive yes-or-no question truthfully with one probability and otherwise supply a random answer. Aggregate frequencies can be estimated from the known randomization procedure.\n\nFor a common binary construction, the randomizer reports the true value with probability\n\n\\[p=\\frac{e^\\varepsilon}{1+e^\\varepsilon}\\]\n\nand the opposite value with probability \\(1-p\\). The ratio between the probabilities of either reported answer under the two possible inputs is bounded by \\(e^\\varepsilon\\). If a fraction \\(\\theta\\) of participants has true value \\(1\\), the expected reported fraction is \\((1-p)+(2p-1)\\theta\\). An estimator can invert this relationship, but the correction amplifies sampling noise, especially when \\(\\varepsilon\\) is small.\n\nThis illustrates the local model’s principal trade-off: weaker trust assumptions usually require substantially more participants to achieve comparable accuracy. Local frequency estimation can perform well at large scale, but sparse categories and high-dimensional data are difficult. Repeated reports by the same person also require composition analysis. Privacy for one report is not automatically privacy for a long-lived stream of reports, particularly when the underlying value changes over time.\n\nIntermediate models distribute trust across several components. In the shuffle model, users randomize messages locally and an intermediary removes their association with senders before aggregation. Under suitable assumptions, shuffling amplifies privacy relative to what the local randomizer alone provides. Secure aggregation, associated with cryptography and distributed computing, can hide individual contributions while revealing an aggregate. It does not by itself make that aggregate differentially private: appropriately calibrated randomness must still be incorporated. Each architecture therefore has a different threat model, including assumptions about collusion, message linkage, compromised participants, and whether the intended noise is actually generated."
    },
    {
      "heading": "Differential privacy in machine learning",
      "text": "A learned model can disclose information about its training data, even when the original examples are not released. Attacks may attempt to determine whether a particular record was used in training, recover distinctive strings, or reconstruct attributes. Overfitting can increase some forms of leakage, but low test error alone is not a privacy guarantee. Differentially private learning instead bounds how the distribution of the trained model changes when a protected training contribution changes. If the entire training procedure is private, publishing the resulting model is covered by that guarantee.\n\nA major technique is differentially private stochastic gradient descent, widely associated with the 2016 work of Martín Abadi and colleagues on deep learning with differential privacy. In a typical version, the algorithm samples a batch, computes a gradient for each example, clips each gradient to a maximum Euclidean norm \\(C\\), adds Gaussian noise to their aggregate, and updates the model. For a per-example gradient \\(g_i\\), clipping takes the form\n\n\\[\\bar g_i = g_i \\min\\!\\left(1,\\frac{C}{\\lVert g_i\\rVert_2}\\right).\\]\n\nClipping bounds each example’s influence, while the injected noise masks that bounded influence. The privacy accountant combines the effects of sampling and repeated updates.\n\nThe exact guarantee depends on details that ordinary deep learning pipelines may treat as implementation choices. Batch sampling must match the accountant’s assumptions. The clipping unit must match the protected unit. If a person contributes many examples, per-example clipping does not automatically yield an equivalent user-level guarantee. Data-dependent preprocessing, vocabulary construction, model selection, and the choice of a clipping threshold can also consume privacy if they use confidential information in ways not already covered by the analysis.\n\nPrivate training changes the optimization problem in several ways. Clipping can bias gradients, while noise increases their variance. Small datasets, rare classes, and high-dimensional parameter spaces can make the resulting accuracy loss substantial. The effects depend on task difficulty, model architecture, public pretraining, the number of optimization steps, and the privacy target. Publicly trained representations followed by private fine-tuning can reduce the amount of private learning required, but the provenance of the supposedly public model matters when interpreting the overall protection.\n\nOther approaches include objective perturbation for suitable empirical risk minimization problems, output perturbation of stable estimators, and teacher-ensemble methods such as PATE. In teacher ensembles, models trained on disjoint private subsets vote on labels, and a noisy aggregation mechanism supplies labels for a student model. In federated learning, client updates may be clipped and noised to protect participants. Federated learning alone is not differential privacy: keeping raw data on a device does not prevent updates from revealing information. Conversely, once a model has been released under a valid private-training guarantee, unlimited processing of that fixed model does not by itself consume additional privacy budget."
    },
    {
      "heading": "Statistical utility and the design of releases",
      "text": "Differential privacy introduces an explicit statistical error source, but the magnitude and consequences depend on the query and population size. A sensitivity-\\(1\\) count released with Laplace noise of scale \\(1/\\varepsilon\\) has absolute error whose distribution does not grow with the population. Its relative error can therefore be small for a large count and severe for a rare event. This asymmetry is central to official statistics and subgroup analysis: a mechanism that performs well on national totals may produce unstable estimates for small geographic areas or uncommon combinations of characteristics.\n\nFor a simple illustration, consider a bounded variable in \\([a,b]\\) and a fixed public sample size \\(n\\), under substitution adjacency. The sensitivity of its sample mean is \\((b-a)/n\\), so Laplace noise can have scale\n\n\\[\\frac{b-a}{n\\varepsilon}.\\]\n\nLarger datasets can therefore support both strong privacy and accurate estimation. However, the bound assumes that the range is genuinely enforced and that the denominator is treated consistently with the privacy model. If the range is chosen from confidential extremes, or if the sample size itself must be protected, additional mechanisms may be required.\n\nError is not limited to the random perturbation added at the last step. Contribution clipping changes the estimand by limiting large contributors. Suppressing or privately selecting categories can omit low-frequency groups. Projecting noisy counts onto nonnegative, integer-valued, or internally consistent tables can introduce bias and correlations. For example, setting all negative noisy counts to zero avoids impossible published values but raises the expected value of counts near zero. These effects matter when users calculate rates, compare groups, or construct uncertainty intervals.\n\nThe best mechanism often depends on the intended query workload rather than on individual queries in isolation. A release containing national, regional, and local counts can exploit the known hierarchical relationships among them. Instead of independently perturbing every requested number, a system may measure a carefully selected set of quantities and derive the rest through mathematical optimization. Accuracy then depends on which measurements were prioritized and on how noise propagates through the reconstruction. Choosing the workload is consequently a substantive part of release design.\n\nPrivately generated synthetic data offers another form of release: artificial records are produced from a private model or private statistical measurements. The guarantee comes from the private construction, not from the fact that records are artificial. A nonprivate generator can memorize real examples. Even a private synthetic dataset may preserve some relationships poorly, especially rare interactions not emphasized during generation. Statistical inference must distinguish sampling uncertainty, privacy noise, model approximation, and post-processing effects; conventional formulas that account only for sampling error may not describe the published data adequately."
    },
    {
      "heading": "Public deployments and engineering requirements",
      "text": "Differential privacy has been deployed in both commercial telemetry and official statistics. Google’s RAPPOR, described in 2014 by Úlfar Erlingsson, Vasyl Pihur, and Aleksandra Korolova, used randomized response and encoding techniques to collect aggregate information about client-side strings. Its design addressed repeated reporting through both persistent and fresh randomization. Apple announced the use of differential privacy in 2016 for learning aggregate usage patterns from devices. Such deployments illustrate the value of collecting population-level signals without treating the collection of exact individual values as a necessary first step.\n\nThe United States Census Bureau used a differentially private disclosure-avoidance system for major products from the 2020 Census. Its TopDown approach allocated accuracy across geographic levels and used optimization to turn noisy measurements into tables and records satisfying selected consistency requirements. The deployment made privacy accounting, small-area accuracy, and the treatment of exact constraints visible to a broad statistical community. Some quantities were designated as invariants and preserved exactly; the privacy interpretation consequently depends on the neighboring-dataset model and the constraints under which the mechanism is analyzed.\n\nPractical database systems face additional difficulties because a query’s sensitivity depends on more than its final expression. Joins, grouping, duplicate records, and repeated events can greatly amplify one person’s influence. A private query service may therefore require explicit user identifiers, contribution bounds, approved aggregation operators, and persistent accounting across sessions. Ordinary SQL syntax does not communicate all of these requirements. A query that counts rows after a many-to-many join can have very different sensitivity from a query that counts distinct protected people.\n\nNumerical implementation is another part of the guarantee. A mechanism proved correct over ideal real numbers can behave differently in finite-precision arithmetic. Early analyses showed that naïve floating-point implementations of noise sampling could leave patterns in the possible outputs that violate intended privacy bounds. Robust implementations use carefully designed discrete mechanisms, secure sampling procedures, or numerical analyses that cover the implemented representation. Discrete Gaussian mechanisms are useful when outputs and computations are naturally integer-valued.\n\nOperational failures can bypass an otherwise valid mechanism. Debug logs may contain raw inputs, a crash report may expose an intermediate unnoised statistic, or an internal dashboard may release a value omitted from the accountant. Randomness must be generated correctly, and private outputs must not be accompanied by data-dependent error messages that escape analysis. Formal verification, specialized type systems, code review, and statistical testing address different parts of this problem. Tests can reveal defects, but observing no successful attack is not a proof of differential privacy; the mathematical claim must cover the actual distribution of all released information."
    },
    {
      "heading": "Interpretation, limitations, and continuing research",
      "text": "The strength of differential privacy is that its guarantee survives arbitrary processing of its outputs and does not require a complete catalog of an attacker’s auxiliary knowledge. Nevertheless, its protection has a precise scope. A private mechanism does not conceal facts already known from another source, erase information from earlier nonprivate releases, or prevent deductions based on genuine population correlations. If an attribute is highly predictable from public information, a private release need not make it unpredictable. The guarantee limits the incremental effect of the protected contribution rather than promising universal ignorance about the participant.\n\nCorrelations between people require particular care. A person’s relative may disclose information relevant to that person, and genetic or household data may contain shared attributes. Standard individual-level differential privacy still bounds the effect of changing the specified individual’s contribution, but it does not automatically protect every fact correlated with that contribution in other people’s records. Group definitions and alternative privacy models can address some broader objectives, usually with increased accuracy costs or additional modeling assumptions. Distinguishing contribution privacy from attribute secrecy prevents claims that exceed what the mathematics establishes.\n\nThere is no universal privacy parameter that makes every application safe. The meaning of a reported \\(\\varepsilon\\) depends on the protected unit, neighboring relation, complete release history, and any accompanying \\(\\delta\\). Different mechanisms can also behave differently while sharing the same worst-case bound. A small number attached to one stage of a pipeline cannot characterize an unrestricted system. Parameter selection therefore combines mathematical accounting with institutional judgments about acceptable disclosure and the value of the statistics being released; the mathematical definition does not settle those judgments by itself.\n\nDifferential privacy is also distinct from legal anonymization, consent, purpose limitation, and data minimization. It supplies evidence about a particular kind of information leakage, not automatic compliance with every requirement of law or professional practice. Information security remains necessary for the unperturbed data and the systems processing it. Similarly, a privately trained model can still produce incorrect, discriminatory, or otherwise problematic outputs. Privacy, predictive accuracy, fairness, and security describe different properties, even when the techniques used to improve them interact.\n\nContinuing research includes tighter privacy accounting, adaptive allocation of privacy budgets, efficient user-level learning, private analysis of graph and longitudinal data, and mechanisms that remain useful for sparse populations. Continual-observation problems are especially challenging because a stream of published totals can repeatedly expose the influence of the same participant. Private graph analysis presents a related distinction between protecting one edge and protecting an entire person’s node and incident edges. Work on machine learning examines how to preserve rare but useful signals without allowing a few training examples to dominate model behavior. Across these problems, the central technical task is to characterize the smallest achievable error under a clearly specified privacy constraint."
    }
  ],
  "links": [
    "randomized-algorithm",
    "data-privacy",
    "computer-science",
    "statistics",
    "machine-learning",
    "encryption",
    "mathematical-function",
    "access-control",
    "cryptography",
    "distributed-computing",
    "training-data",
    "overfitting",
    "deep-learning",
    "variance",
    "fine-tuning",
    "empirical-risk-minimization",
    "federated-learning",
    "mathematical-optimization",
    "synthetic-data",
    "population-census",
    "database",
    "sql",
    "formal-verification",
    "law",
    "information-security"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}