{"id":1130,"date":"2026-07-04T06:03:36","date_gmt":"2026-07-03T23:03:36","guid":{"rendered":"https:\/\/sumberlaba.com\/index.php\/2026\/07\/04\/what-is-differential-privacy-a-comprehensive-guide-to-understanding-and-implementing-differential-privacy\/"},"modified":"2026-07-04T06:03:36","modified_gmt":"2026-07-03T23:03:36","slug":"what-is-differential-privacy-a-comprehensive-guide-to-understanding-and-implementing-differential-privacy","status":"publish","type":"post","link":"https:\/\/sumberlaba.com\/index.php\/2026\/07\/04\/what-is-differential-privacy-a-comprehensive-guide-to-understanding-and-implementing-differential-privacy\/","title":{"rendered":"What is Differential Privacy? A Comprehensive Guide to Understanding and Implementing Differential Privacy"},"content":{"rendered":"<h1>What is Differential Privacy? A Comprehensive Guide to Understanding and Implementing Differential Privacy<\/h1>\n<p>In an era where data is often called the \u201cnew oil,\u201d organizations collect vast amounts of personal information to drive decision-making, improve services, and unlock insights. However, this treasure trove of data comes with a dark side: the risk of exposing sensitive details about individuals. Traditional anonymization techniques, such as removing names and addresses, have proven woefully inadequate\u2014adversaries can often re-identify individuals by linking seemingly harmless data with other public sources. This is where differential privacy enters as a game-changer. It is a mathematically rigorous framework that allows analysts to extract aggregate statistics from a dataset while providing a formal, quantifiable guarantee that the privacy of any individual in the dataset is preserved. In essence, differential privacy ensures that the output of a data analysis does not reveal whether any particular person\u2019s information was included or not. This gives individuals plausible deniability, making it nearly impossible for an attacker to infer their presence or their specific data values, regardless of any auxiliary information the attacker may possess.<\/p>\n<p>The roots of differential privacy can be traced back to the early 2000s, primarily through the work of Cynthia Dwork, Frank McSherry, Kobbi Nissim, and Adam Smith. Their seminal 2006 paper, \u201cCalibrating Noise to Sensitivity in Private Data Analysis,\u201d laid the foundation for what we now know as differential privacy. Since then, it has been adopted by major tech companies like Google (for RAPPOR in Chrome), Apple (for collecting user usage patterns), and Microsoft (for telemetry systems). Perhaps its highest-profile deployment was by the U.S. Census Bureau for the 2020 Decennial Census, where differential privacy was used to protect the privacy of every American while still publishing essential demographic and redistricting data. The core idea is elegant yet powerful: instead of removing identifying features, we add a calibrated amount of randomness to the query results. This noise masks the contribution of any single individual, thereby preserving privacy without rendering the data useless. The challenge, of course, lies in balancing the level of privacy (controlled by a parameter called epsilon) with the utility of the data.<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/via.placeholder.com\/800x600\/4a90d9\/ffffff?text=what%20is%20differential%20privacy\" alt=\"Article illustration\" style=\"display:block;margin:20px auto;max-width:100%;height:auto;border-radius:8px;\" \/><\/p>\n<h2>Step-by-Step Guide to Understanding Differential Privacy<\/h2>\n<h3>Step 1: Grasping the Core Intuition \u2013 You Are Not the Data Point<\/h3>\n<p>To truly understand differential privacy, you must first internalize the concept of \u201cplausible deniability.\u201d Imagine you are in a large room with hundreds of people, and a researcher asks everyone a sensitive question, like \u201cDo you support a controversial policy?\u201d The researcher does not want to know individual answers; they only want the overall percentage. In a differential privacy scenario, the researcher does not look at your specific answer. Instead, they design a mechanism that takes everyone\u2019s answers and intentionally scrambles them in a controlled way. For example, the mechanism might flip a coin: if the coin comes up heads, you answer truthfully; if tails, you answer randomly (say, \u201cyes\u201d or \u201cno\u201d with equal probability). Now, even if the researcher sees a \u201cyes,\u201d they cannot be sure whether it came from a true \u201cyes\u201d or from the random noise. Your privacy is protected because the output of the mechanism is roughly the same whether you are in the dataset or not. This is the fundamental promise of differential privacy: the outcome of an analysis is almost indistinguishable between two datasets that differ by only one person\u2019s data. The strength of this guarantee is quantified by a parameter \u03b5 (epsilon), where a smaller epsilon means stronger privacy (outputs are more similar) and a larger epsilon means weaker privacy but often higher accuracy.<\/p>\n<h3>Step 2: Understanding the Formal Definition and Epsilon<\/h3>\n<p>Let\u2019s turn the intuition into a mathematical definition. A randomized algorithm M (which takes a dataset D and returns some output) is said to be \u03b5-differentially private if for any two neighboring datasets D and D&#8217; (datasets differing by exactly one individual\u2019s data) and for any set of possible outcomes S \u2286 Range(M), the following holds:<\/p>\n<p style=\"text-align:center;\">Pr[M(D) \u2208 S] \u2264 e^\u03b5 \u00d7 Pr[M(D&#8217;) \u2208 S]<\/p>\n<p>In simpler terms, the probability of getting a particular output does not change by more than a multiplicative factor of e^\u03b5 when one person\u2019s data is added or removed. If \u03b5 = 0, the ratio is 1, meaning the output distributions are identical, and perfect privacy is achieved\u2014but the result would be completely independent of the data, hence useless. Realistic values of \u03b5 lie between 0.1 and 10, depending on the use case. For extremely sensitive data (e.g., medical records), researchers often aim for \u03b5 \u2264 1. For less sensitive or aggregated telemetry, \u03b5 can be higher. The term e^\u03b5 is called the privacy loss. If \u03b5 = 0.1, e^\u03b5 \u2248 1.105, meaning the output probabilities change by at most 10.5% between two neighboring datasets. This is a strong guarantee. If \u03b5 = 1, the change is about 2.718 times, which is still meaningful but offers less protection. It is crucial to understand that \u03b5 is not a \u201cprobability\u201d but a \u201cbudget\u201d that gets consumed. Every query on the same dataset using differential privacy costs some epsilon, and the total epsilon spent over a series of queries must be accounted for (this is the composition property we\u2019ll discuss later).<\/p>\n<h3>Step 3: Learning the Core Mechanisms \u2013 Laplace, Gaussian, and Exponential<\/h3>\n<p>Implementing differential privacy requires adding noise to the output of a function. The most common mechanism for numeric queries (e.g., count, sum, mean) is the Laplace mechanism. It adds noise drawn from a Laplace distribution (a double exponential distribution) centered at zero with a scale parameter that depends on the sensitivity of the query and the desired \u03b5. Sensitivity is the maximum possible change in the output when one person\u2019s data is added or removed. For a count query, the sensitivity is 1 (because adding one person can change the count by at most 1). The Laplace mechanism adds noise Lap(\u0394f \/ \u03b5), where \u0394f is the sensitivity. So, if you want \u03b5 = 0.5 for a count, you add noise with scale 1\/0.5 = 2. This means the noise is fairly large but still allows a reasonable estimate for large counts. Another mechanism is the Gaussian mechanism, which adds noise from a Gaussian (normal) distribution. However, the Gaussian mechanism only satisfies (\u03b5, \u03b4)-differential privacy, a relaxed version where \u03b4 is a small probability of failure. This is often used when you need to output sums of high-dimensional vectors (like in deep learning with DP-SGD). The third key mechanism is the Exponential mechanism, which is used for non-numeric outputs (e.g., selecting the most frequent category). It assigns a score to each possible output and selects one with probability proportional to the exponential of (score \u00d7 \u03b5 \/ (2\u0394score)). This mechanism is powerful for tasks like releasing a histogram or a decision tree.<\/p>\n<table border=\"1\" cellpadding=\"8\" cellspacing=\"0\" style=\"border-collapse:collapse; width:100%; margin:20px 0;\">\n<caption>Table 1: Comparison of Common Differential Privacy Mechanisms<\/caption>\n<thead>\n<tr style=\"background-color:#f2f2f2;\">\n<th>Mechanism<\/th>\n<th>Type of Output<\/th>\n<th>Noise Distribution<\/th>\n<th>Privacy Guarantee<\/th>\n<th>Key Parameter<\/th>\n<th>Typical Use Cases<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Laplace<\/td>\n<td>Numeric (real-valued)<\/td>\n<td>Laplace(0, b)<\/td>\n<td>\u03b5-DP (strict)<\/td>\n<td>scale b = \u0394f \/ \u03b5<\/td>\n<td>Count queries, sums, means<\/td>\n<\/tr>\n<tr>\n<td>Gaussian<\/td>\n<td>Numeric (real-valued)<\/td>\n<td>Gaussian(0, \u03c3\u00b2)<\/td>\n<td>(\u03b5, \u03b4)-DP (approximate)<\/td>\n<td>\u03c3 = \u0394f * sqrt(2 ln(1.25\/\u03b4)) \/ \u03b5<\/td>\n<td>High-dimensional queries, gradient noise in ML<\/td>\n<\/tr>\n<tr>\n<td>Exponential<\/td>\n<td>Non-numeric (e.g., category, structure)<\/td>\n<td>Probability distribution over outcomes<\/td>\n<td>\u03b5-DP (strict)<\/td>\n<td>Score function with sensitivity \u0394<\/td>\n<td>Top-k selection, histogram bin choice, model selection<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h3>Step 4: Mastering Composition Properties \u2013 Managing the Privacy Budget<\/h3>\n<p>In real-world scenarios, you rarely run just one query. You might ask multiple questions about the same dataset\u2014for example, \u201cHow many users are over 30?\u201d and \u201cWhat is the average income?\u201d If each query is differentially private in isolation, the combined result might leak more privacy than intended. This is where composition theorems become critical. The simplest and most intuitive is the sequential composition theorem: if you run k mechanisms, each satisfying \u03b5_i-differential privacy (and each potentially depending on the same dataset), then the whole process satisfies (\u2211 \u03b5_i)-differential privacy. In other words, the privacy budgets add up. So if you plan to ask ten queries with \u03b5 = 0.5 each, the total \u03b5 is 5, which may be too high for sensitive data. You must budget accordingly. The parallel composition theorem is much more friendly: if you partition the dataset into disjoint subsets and run a DP mechanism on each subset independently, the total privacy guarantee is only the maximum \u03b5 used on any subset, not the sum. This is because each person\u2019s data appears in only one subset. So, to maximize utility under a fixed budget, you often design queries on disjoint subsets of data. There are also advanced composition theorems (e.g., the moment accountant used in DP-SGD) that allow tighter accounting than simple summation, especially when \u03b5 is small. These are more complex to understand but essential for efficient deep learning applications.<\/p>\n<h3>Step 5: Exploring Real-World Deployments \u2013 From Google to the U.S. Census<\/h3>\n<p>To see differential privacy in action, look at how Apple uses it to collect emoji usage statistics and typing patterns. Apple\u2019s implementation locally adds noise on the user\u2019s device before sending the data to Apple\u2019s servers. This is called local differential privacy (LDP), where the trusted curator does not exist, and each user applies the DP mechanism to their own data. This gives the user strong privacy guarantees even from the data collector. Apple uses a privacy budget of around \u03b5 = 4 for their count queries, which was controversial (some argued it was too high), but it demonstrated a massive real-world adoption. Google\u2019s RAPPOR system is another famous example; it uses a randomized response technique combined with Bloom filters to collect browser statistics (e.g., blocked extensions) from Chrome users without revealing which specific extensions are installed by which user. The U.S. Census Bureau\u2019s use of differential privacy for the 2020 census is perhaps the most rigorous: all published tables (age, race, housing) were produced using a differentially private algorithm called the TopDown algorithm. The total privacy budget was split across geographic hierarchies, with a total \u03b5 of about 5.46 across all tables. This required a massive effort to tune parameters and ensure that the resulting data remained useful for redistricting. The Census release included significant privacy loss budget tables for each type of query, allowing data users to understand the trade-offs.<\/p>\n<table border=\"1\" cellpadding=\"8\" cellspacing=\"0\" style=\"border-collapse:collapse; width:100%; margin:20px 0;\">\n<caption>Table 2: Typical Epsilon Values in Practice and Their Implications<\/caption>\n<thead>\n<tr style=\"background-color:#f2f2f2;\">\n<th>Epsilon Range<\/th>\n<th>Privacy Level<\/th>\n<th>Noise Magnitude (for count of 1M)<\/th>\n<th>Example Use Case<\/th>\n<th>Risk to Individuals<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>0.01 \u2013 0.1<\/td>\n<td>Very high<\/td>\n<td>Very large (noise > 1000)<\/td>\n<td>Highly sensitive medical research<\/td>\n<td>Extremely low; outputs barely change<\/td>\n<\/tr>\n<tr>\n<td>0.1 \u2013 1.0<\/td>\n<td>High<\/td>\n<td>Large (noise ~100-1000)<\/td>\n<td>U.S. Census 2020 (some queries), academic studies<\/td>\n<td>Very unlikely to re-identify individuals<\/td>\n<\/tr>\n<tr>\n<td>1.0 \u2013 4.0<\/td>\n<td>Moderate<\/td>\n<td>Moderate (noise ~10-100)<\/td>\n<td>Apple\u2019s emoji collection, Google RAPPOR<\/td>\n<td>Some theoretical risk if adversary has strong prior knowledge<\/td>\n<\/tr>\n<tr>\n<td>4.0 \u2013 10.0<\/td>\n<td>Low<\/td>\n<td>Small (noise < 10)<\/td>\n<td>Internal analytics without high sensitivity<\/td>\n<td>Real risk of inference if large dataset<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h2>Tips and Best Practices for Implementing Differential Privacy<\/h2>\n<h3>Tip 1: Choose Epsilon Based on Both Utility and Context, Not Just an Arbitrary Number<\/h3>\n<p>One of the most common mistakes newcomers make is treating \u03b5 as a binary switch (\u201cprivate enough\u201d vs \u201cnot\u201d). In reality, \u03b5 must be chosen carefully by considering the sensitivity of the data, the trust relationship with users, and the required accuracy of the analysis. There is no universal \u201csafe\u201d \u03b5; a value that works for a public opinion poll (\u03b5 = 2) might be dangerously high for health records. Always start with a literature review for similar applications. For instance, many academic papers recommend \u03b5 = 1 as a starting point. However, if you are deploying differential privacy for a consumer product, you should also consider transparency: tell your users what \u03b5 you use and why. Apple initially used \u03b5 = 4 for emoji data, which raised some eyebrows, but they later reduced it for more sensitive categories. Furthermore, if your dataset is small (e.g., 100 people), a moderate \u03b5 like 1 may still add too much noise to be useful. In such cases, you may need to accept a higher \u03b5 or consider alternative approaches like synthetic data generation with DP. The golden rule is to simulate: run your analysis with and without noise on a sample of real data (or synthetic data) to see the accuracy loss for your specific metrics.<\/p>\n<h3>Tip 2: Use Established Libraries and Frameworks Instead of Building from Scratch<\/h3>\n<p>Differential privacy is notoriously tricky to implement correctly. Even a small mistake in handling sensitivities, composition, or random number generation can break the guarantee. Therefore, whenever possible, rely on well-audited open-source libraries. For Python, the most popular is <strong>Google&#8217;s differential-privacy library<\/strong> (now part of the OpenDP project). It provides implementations of Laplace, Gaussian, and exponential mechanisms, as well as building blocks for sums, means, histograms, and count queries. There is also <strong>IBM\u2019s diffprivlib<\/strong>, which integrates seamlessly with scikit-learn for machine learning tasks. For R, the <strong>diffpriv<\/strong> package is available on CRAN. For deep learning, <strong>PyTorch\u2019s Opacus<\/strong> and <strong>TensorFlow Privacy<\/strong> implement differentially private stochastic gradient descent (DP-SGD). When using these libraries, you still need to set epsilon and delta, but the complicated mathematics is handled for you. Always read the documentation thoroughly; most libraries assume you know the sensitivity of your query. For example, if you query a mean, you must provide the bounds of the data (e.g., income between 0 and 1,000,000), which determines the sensitivity. Never skip this step because an unbounded query can have infinite sensitivity, and the library might not warn you adequately.<\/p>\n<h3>Tip 3: Plan for Privacy Budget Allocation Across Multiple Queries and Workflows<\/h3>\n<p>In a production system, you often need to answer many queries over time from the same dataset. You must manage a total privacy budget (epsilon) that gets consumed. This requires careful design of your data release strategy. One best practice is to think of a \u201cprivacy budget dashboard\u201d where each query type is assigned a portion of the total budget. For example, you might allocate 40% of your total \u03b5 to overarching demographic statistics, 30% to behavioral summaries, and 30% to exploratory analysis. If you exceed the budget, you cannot run more queries on that dataset (unless you add more noise to compensate, but that would change the already released outputs). Another technique is to use adaptive composition with a \u201cprivacy accountant\u201d that tracks the cumulative privacy cost in a tighter way than simple summing. Libraries like the Google DP library include an accountant that implements the moment accountant, which gives a much better bound than the na\u00efve sum for small \u03b5. Also, consider using <strong>subsampling amplification<\/strong>: if you only run queries on a random subset of the data (e.g., 10% of users), the privacy guarantee improves (the effective \u03b5 is multiplied by the subsampling ratio). Many modern DP implementations, especially in DP-SGD, automatically account for this amplification.<\/p>\n<h2>Frequently Asked Questions About Differential Privacy<\/h2>\n<h3>Q1: Does a larger epsilon mean worse privacy?<\/h3>\n<p>Yes, absolutely. Epsilon is the main measure of privacy loss. A smaller epsilon means that the outputs of the mechanism on two neighboring datasets are more similar, which gives stronger privacy. For \u03b5 = 0, the outputs are identical regardless of whether any individual is in the dataset\u2014perfect privacy but zero utility. For \u03b5 = 10, the multiplicative factor e^10 \u2248 22,026, meaning the probabilities can change enormously, offering very weak privacy. In practice, you must choose an epsilon that balances the need for utility against the acceptable privacy risk. For many applications, values below 1 are considered strong, while values above 5 are considered weak. The U.S. Census Bureau used a total budget around 5.46, which was controversial but deemed necessary for releasing enough accurate data for redistricting.<\/p>\n<h3>Q2: Is differential privacy the same as anonymization?<\/h3>\n<p>No, it is far more robust. Traditional anonymization techniques, such as removing names, IDs, ZIP codes, and dates of birth, rely on heuristic rules. They often fail because an adversary can link residual attributes (like gender, age, and job title) with other public datasets to re-identify individuals\u2014a known risk demonstrated in the Netflix Prize and AOL search data releases. Differential privacy provides a mathematical guarantee that the presence or absence of a single individual cannot be inferred with significant confidence, regardless of auxiliary information. It does not rely on removing attributes; it adds noise. So even if an attacker has complete knowledge of all other records, they cannot be sure about the individual in question. That said, differential privacy is not a silver bullet: it protects the individual\u2019s participation and data value, but it cannot protect against all inferences (e.g., known aggregate statistics about a group that includes the individual).<\/p>\n<h3>Q3: Can differential privacy be used for small datasets?<\/h3>\n<p>It can, but with significant challenges. Differential privacy noise scales inversely with the privacy budget and proportionally with the query sensitivity. For small datasets (e.g., 100 people), even a moderate epsilon like 1 can add so much noise that the query result becomes meaningless. For example, a count of 50 with added Laplace noise of scale 1 might give a value between 20 and 80, which is very imprecise. However, if your query has low sensitivity (like a count per cell in a histogram with small bins), you can still get useful results if the statistical signal is strong relative to the noise. In practice, differential privacy is more suitable for large datasets (at least several thousand records) where the noise is small relative to the aggregate value. For small datasets, alternative privacy models like k-anonymity or synthetic data generation may be considered, but they lack the formal guarantees of DP.<\/p>\n<h3>Q4: What is the difference between local and central differential privacy?<\/h3>\n<p>In central DP, there is a trusted curator who holds the entire original dataset (e.g., a hospital\u2019s patient database). The curator applies the DP mechanism (adds noise) before releasing the answers to analysts, and the noise is calibrated for the whole dataset. This usually allows higher accuracy because the noise required for a single query is smaller (the sensitivity depends on the global query, not on individual contributions). In local DP, there is no trusted curator. Each individual applies the DP mechanism to their own data on their own device (e.g., Apple collecting emoji usage). The collector receives already-noised data. Because each individual adds noise independently, the total noise grows faster (the variance adds up across users), so local DP typically requires many more users to achieve the same accuracy as central DP for the same epsilon. However, local DP offers stronger trust guarantees because the data collector never sees the raw data. The choice between the two depends on the threat model and the availability of a trusted aggregator.<\/p>\n<h3>Q5: How do you explain differential privacy to non-experts (e.g., executives or users)?<\/h3>\n<p>It can be challenging, but an effective analogy is the \u201crandomized response\u201d coin flip example. Say something like: \u201cImagine you want to know how many people in a group have a certain trait, but you don\u2019t want to know who does. Differential privacy is like asking each person to flip a coin. If the coin lands heads, they answer truthfully; if tails, they answer randomly. The overall percentage is corrected statistically, but even if I see a \u2018yes\u2019 I cannot be sure if it came from a true yes or the random flip. That way, you are never certain about any individual. The level of uncertainty is controlled by a number called epsilon, and we choose it to be strong enough to protect your privacy but not so strong that the overall statistics become useless.\u201d Avoid using mathematical formulas. Emphasize the guarantee that the results \u201clook pretty much the same whether you are in the dataset or not.\u201d Also, mention that major companies and governments use it to prove they care about your privacy while still learning from data.<\/p>\n<h2>Conclusion<\/h2>\n<p>Differential privacy represents a paradigm shift in how we think about data protection. Instead of relying on ad-hoc removal of identifiers, it provides a rigorous framework that can be proved mathematically to limit privacy risk. By adding controlled noise calibrated to the sensitivity of queries, differential privacy ensures that the output of any analysis does not reveal whether any specific individual\u2019s data was included. This empowers organizations to share valuable statistics without exposing personal information, and it gives individuals a verifiable guarantee that their data may be used but not abused. Throughout this guide, we have explored the core intuition, the formal definition, the primary mechanisms (Laplace, Gaussian, and Exponential), the critical concept of composition, and real-world deployments that demonstrate its feasibility. We have also discussed best practices for choosing privacy parameters, using trusted libraries, and budgeting epsilon across multiple analyses. The FAQ covered common questions to help you clarify misconceptions and communicate differential privacy to diverse audiences.<\/p>\n<p>As data regulations continue to tighten and privacy expectations from users increase, differential privacy is no longer just an academic curiosity\u2014it is becoming an essential tool in the data scientist\u2019s toolkit. Whether you are analyzing medical records, improving a recommendation system, or providing open government data, integrating differential privacy into your workflow will help you build trust, comply with privacy laws like GDPR and CCPA, and protect the individuals behind the data. The journey to adopt differential privacy can be challenging, but the rewards are immense: you can continue to derive insights from data without sacrificing the privacy of the people you serve. Start small, use the available libraries, and always simulate the trade-off between utility and privacy. With careful design, you can contribute to a future where data-driven innovation and individual privacy coexist harmoniously.<\/p>\n","protected":false},"excerpt":{"rendered":"<p>What is Differential Privacy? A Comprehensive Guide to Understanding and Implementing Differential Privacy In an era where data is often called the \u201cnew oil,\u201d organizations collect vast amounts of personal information to drive decision-making, improve services, and unlock insights. However, this treasure trove of data comes with a dark side: the risk of exposing sensitive &hellip; <\/p>\n","protected":false},"author":2716,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"om_disable_all_campaigns":false,"_monsterinsights_skip_tracking":false,"footnotes":""},"categories":[],"tags":[],"class_list":["post-1130","post","type-post","status-publish","format-standard","hentry"],"aioseo_notices":[],"_links":{"self":[{"href":"https:\/\/sumberlaba.com\/index.php\/wp-json\/wp\/v2\/posts\/1130","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/sumberlaba.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/sumberlaba.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/sumberlaba.com\/index.php\/wp-json\/wp\/v2\/users\/2716"}],"replies":[{"embeddable":true,"href":"https:\/\/sumberlaba.com\/index.php\/wp-json\/wp\/v2\/comments?post=1130"}],"version-history":[{"count":0,"href":"https:\/\/sumberlaba.com\/index.php\/wp-json\/wp\/v2\/posts\/1130\/revisions"}],"wp:attachment":[{"href":"https:\/\/sumberlaba.com\/index.php\/wp-json\/wp\/v2\/media?parent=1130"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/sumberlaba.com\/index.php\/wp-json\/wp\/v2\/categories?post=1130"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/sumberlaba.com\/index.php\/wp-json\/wp\/v2\/tags?post=1130"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}