AI-based context-aware data anonymization with federated adaptive differential privacy for IoT applications

Abstract

The rapid growth of Internet of Things (IoT)-based healthcare systems has raised significant concerns regarding data privacy and security. Ensuring privacy while maintaining the utility of healthcare data remains a major challenge in IoT-enabled healthcare environments. This article proposes a novel data anonymization framework based on federated learning and adaptive differential privacy. Initially, IoT healthcare data are processed using a residual bidirectional gated recurrent unit (Res-BiGRU) model to capture contextual and privacy-related features. Subsequently, an adaptive differential privacy mechanism is applied to minimize information loss while ensuring strong privacy protection. To improve optimization performance, a modified Resilient Adaptive Apiary Organizational Optimization Algorithm (RAOOA) incorporating a fitness-based adaptive factor is introduced to enhance convergence stability and solution quality. The effectiveness of the proposed framework is evaluated using a healthcare dataset. The proposed approach demonstrates superior performance compared with existing methods in terms of computation time, information loss, and privacy risk. The results indicate that the proposed framework provides an efficient and scalable solution for privacy-preserving IoT healthcare data management while achieving an improved balance between data utility, computational efficiency, and privacy protection.

1 Introduction

The Internet of Things (IoT) has transformed the way people communicate with the digital world by linking a huge array of systems, varying from smart home applications to industrial sensors. Although this interlinked ecosystem provides more efficiency and convenience, it also raises privacy concerns (Murakami and Sei, 2023). The simultaneous acquisition, transmission, and execution of data in IoT devices provide specific privacy risks and vulnerabilities for organizations and users (Ren et al., 2024). The open features of IoT environments increase the privacy concerns as “things” are autonomous with a certain degree of authority to distribute their knowledge and capabilities to satisfy their collective or individual operations (Wang et al., 2022). The privacy of the IoT is a unique consideration needed to secure the individual’s data from the IoT environment’s exposure (Zhu et al., 2022). These phases are significant because an IoT setup allows almost any logical or physical domain or object to be provided a unique identity and the capability to interact autonomously over a similar network or the Internet. This seeks to make the IoT experiment simple by bringing together, contrasting, and comparing disparate experimental contributions (Sun et al., 2023).

Healthcare devices across the world are transforming digital records, enhancing medical research, revolutionizing patient care, and improving operational efficiency (Zhao et al., 2023). The transformation toward telemedicine platforms, electronic health records (EHRs), and cloud-assisted healthcare programs has greatly improved the accessibility of the data (Nawshin et al., 2024), coordination between healthcare experts, and overall service delivery. However, this digital revolution comes with specific cybersecurity complexities, as confidential patient data become a profitable target for cybercriminals (Wu et al., 2023). The healthcare domain has faced an increase in cyberattacks, including insider threats, phishing methods, data breaches, and ransomware attacks. These security cases not only compromise the privacy of the patient but also affect healthcare tasks, threatening lives (Salim et al., 2022). To resolve these concerns, robust anonymization approaches must be developed to safeguard the healthcare data pipelines from cyber exploitation and unauthorized access (Devineni and Gorantla, 2023).

Privacy has become a significant concern. The term privacy defines the personal data about an individual that cannot be revealed to others. Personal data disclosure may impact the individual financially, psychologically, and emotionally (Liu et al., 2022). For instance, if an individual has cancer, disclosing the data to their employer or insurance company may impact their livelihood. Therefore, it is significant to eliminate privacy disclosure at the data publication (Sfar et al., 2022) and collection time (Goel and Pujari, 2023). However, the privacy-preserving data acquisition and publication are complex, as they highly depend on the data protection policy. Anonymization on the basis of privacy protection approaches has become important (Pasham, 2023). The anonymity approach preserves sensitive data’s privacy via suppression and generalization approaches (Fernández et al., 2022). However, the conventional anonymity approaches, such as t-closeness, L-diversity, and k-anonymity, are susceptible to numerous privacy concerns, such as skewness attacks, similarity concerns, sensitive attribute concerns, homogeneity attacks, attribute disclosure, background knowledge attacks, membership, and identity disclosure (Jayapradha et al., 2022). The privacy-preserving data collection approach for healthcare IoT service models varies from normal methods. Unlike static data, the IoT framework ensures data that evolve with time (referred to as a “data stream”). Hence, it is significant to design an effective privacy-preserving approach for the healthcare IoT (Xie et al., 2023). Therefore, an effective model is designed in this work to minimize the risk of privacy concerns while maintaining the data utility by minimizing the privacy utility balance for the healthcare sector.

While the proposed approach mainly considers federated learning and adaptive differential privacy in order to anonymize IoT data, it can be applied to the infrastructure that utilizes blockchain technology. Blockchain technology offers decentralized trust, immutability, and auditability of the anonymization process. In this case, federated learning model updates, privacy budgets, and anonymization decisions can be stored in a distributed ledger. This will help increase the transparency of the system and thus promote explainable AI because the process of data transformation and decision-making will be traceable.

The structure of the designed data anonymization framework is elucidated here. The conventional data anonymization works are presented in Section 2, and the new concept of AI-based data anonymization in the IoT healthcare domain is described in Section 3. The recommendations of Res-BiGRU for data context and privacy level prediction and a revamped algorithm for optimization are given in Section 4. The proposal of FL-ADP for data anonymization in IoT healthcare is shown in Section 5, and the experiments for this framework are also provided. The conclusion is finally given in Section 6.

2 Related works

In 2024, Kil et al. (2024) implemented optimal feature selection approaches that result in increased utility and privacy by tuning the utility-privacy balance for resource-limited system environments. This approach has retrieved the privacy-aware features by choosing features demanding privacy protection and eliminating privacy-sensitive features, making it complex for attackers to recognize the data. The results have illustrated that the model enhanced data privacy.

Irshad et al. (2023) designed a smart approach for the smart cities IoT sector that concentrated on improving communication reliability and efficiency, while reducing latency concerning data security and energy consumption for data transmission. This approach leveraged an optimization algorithm to improve the security, efficiency, and reliability of data transmission in IoT-based networks. The model minimized the latency and increased the reliability.

Denis et al. (2023) presented a decentralized usage control approach to improve IoT systems to handle the data they produce. The work considered the usage control system as a distributed ledger component within the network. The results of the experiments on the private network have minimized the time.

Chhetri et al. (2024) introduced a novel technology-aided model that has allowed data distribution while guaranteeing that the distributed data were interoperable and that the included required context data were trustworthy and accurate. The observations illustrated that the model could ensure both data veracity and quality. Chen et al. (2023) developed an FL-based model with effective performance. This model secured privacy with low overhead. Further, this work integrated the FL with blockchain. The experiments showcased the practical performance of the designed system.

Alhija et al. (2024) discussed the primary trade-off aspects, such as the range of cryptographic approaches and consensus protocols. This approach offered actionable suggestions to healthcare sectors to support them in benefiting from the vast promise of the blockchain for linked IoT healthcare by providing a trade-off among decentralization performance and merits, compliance, and security demands.

Liu et al. (2024) presented a protection solution for data privacy and integrity. Most importantly, this model secured the integrity of the data via the distributed devices and blockchain mechanism. The model concentrated on maintaining the data usability for further analytical operations of the data users. This model efficiently encrypted the source IoT data based on the data users’ demands, while the data integrity could be secured under the blockchain.

Kumar et al. (2023) designed a privacy preservation approach in the Industrial Internet of Things (IIoT) employing the improvements of AI methods. The two primary phases of the implemented model were the restoration and the sanitization of the IIoT data. The simulation results show the dominance of the designed framework. This model achieved high privacy preservation.

2.1 Research gaps and challenges

2.2 Integration with blockchain for security and explainability

The integration of blockchain into the proposed framework would ensure tamper-proof and secure storage of IoT healthcare data. The reason is that blockchain creates an environment that does not depend on any one entity controlling the data, thus ensuring trust among its participants. In this new architecture, data would be generated through the IoT devices and be analyzed using the federated learning method. The data would then undergo adaptive differential privacy-based anonymization locally. Model updates and privacy parameters obtained during the process would be stored on the blockchain ledger, while smart contracts would be used to control privacy and access constraints and ensure secure automation. Blockchain would enhance explainability because it ensures transparency and traceability of the processes involving model updates and anonymization. Each transaction conducted through the blockchain would allow interested stakeholders to track how the data were processed, thus increasing the level of accountability in the system. Thus, the proposed RAOOA-FL-ADP framework would successfully integrate into blockchain-based systems.

The optimization techniques used by Falahah et al. (2024), Givi and Hubalovska (2023), and Amiri et al. (2024) have been chosen as baselines for comparisons due to their effectiveness in optimization processes, robustness in convergence processes, and previous implementations in privacy-related and data-driven applications.

3 A new concept of AI-based data anonymization in the IoT healthcare domain

3.1 Anonymization in the IoT-based healthcare domain

Anonymization in the IoT-based healthcare sector is a significant practice to secure patient privacy while still allowing the high potential of data-driven technologies in healthcare services. The high growth of the IoT in the healthcare sector via systems such as smart diagnostic tools, remote monitoring systems, wearable fitness trackers, and implantable sensors has resulted in the continuous production of a huge amount of confidential health data (Andrew et al., 2023). Mostly, these data include personally identifiable information (PII) such as biometric readings, medical conditions, addresses, names, and even location data, all of which, if revealed, could compromise the confidentiality of the data and result in significant legal and ethical problems. Anonymization plays a significant part by transforming certain identifiable data in a way that eliminates the recognition of individuals, even when the data sources are integrated with external data. Unlike conventional models, IoT environments pose unique complexities because of their distributed model, real-time data flow, constrained device computational power, and data type heterogeneity. Efficient anonymization in this context demands not only potential privacy methods but also a careful trade-off to maintain the utility of the data for clinical diagnosis, decision-making, and predictive analytics. Overly aggressive anonymization processes can lead to high data loss, undermining the reliability and the quality of insights extracted from the data. In contrast, weak anonymization leaves the data susceptible to redetection attacks (Prabha and Chatterjee, 2024).

The improved anonymization approaches, such as L-diversity, pseudonymization, generalization, differential privacy, and suppression, are adaptive to manage the particular demands of the IoT healthcare models. As the sector evolves, smart privacy-preserving approaches that incorporate context-aware and adaptive anonymization are becoming important to guarantee that sensitive patient data are secured, without compromising the usefulness and integrity of the data. Figure 1 presents the diagrammatic form of anonymization in the IoT-based healthcare domain.

FIGURE 1

3.2 Overall concept of proposed data anonymization framework in the IoT healthcare domain

In the IoT healthcare sector, data anonymization is a crucial task to secure the sensitive data of patients while allowing its usage for clinical monitoring, analysis, and decision-making. Conventional anonymization approaches face some complexities that limit their efficiency in the sensitive and dynamic environment. The conventional anonymization approaches, such as data masking, L-diversity, and k-anonymity, often struggle to maintain the balance between data utility and privacy, most importantly when employed on real-time, high-dimensional IoT data. These approaches may lead to significant data loss, minimizing the accuracy of the healthcare analytics and affecting early healthcare decisions. In addition, most of the existing methods consider a centralized data environment, which is infeasible in IoT-based healthcare devices, where the data are produced and processed across distributed systems with constrained computational resources.

This centralized mechanism raises concerns about data breaches and single points of failure. In addition, the static privacy methods employed in conventional approaches are not well-suited for adaptive privacy demands, where the data sensitivity may change across time, contexts, and patients. As a result, the fixed anonymization approaches can either over-protect, resulting in unusable data, or under-protect, leaving patients susceptible to re-identification. These problems highlight the requirement for a context-aware, intelligent, and adaptive data anonymization model that can operate effectively in resource-limited, decentralized IoT healthcare environments while preserving both practical utility and the privacy of the data. Therefore, a new data anonymization framework is introduced for the IoT healthcare domain. Its pictorial representation is given in Figure 2.

FIGURE 2

The suggested data anonymization system in the IoT healthcare sector presents an innovative strategy to guarantee strong privacy protection while preserving the sensitive medical data’s utility. In this strategy, the data collection is initially performed from the available datasets. In the further stage, the Res-BiGRU technique is utilized to validate the context and sequential nature of the data. This deep learning approach improves the capability of the system to understand the temporal dependencies and estimate certain privacy levels for distinct data attributes, which is significant for context-aware anonymization. This model assesses the data sensitivity, allowing for tailored anonymization methods rather than a one-size-fits-all mechanism. After this, the FL-ADP is employed to perform privacy-preserving data processing across the decentralized IoT systems without distributing the original data, minimizing the concern of data breaches. This model guarantees dynamic privacy adjustments on the basis of data sensitivity; especially, FL enhances the accuracy of the approach by learning from distinct resources collaboratively. The merit of this stage is a privacy-preserving, decentralized learning approach that maintains high model performance. To improve the anonymization task, the RAOOA is employed to optimize the trade-off between information loss and data utility. This approach fine-tunes the privacy parameters to guarantee less information loss, which is crucial in the healthcare sector, where real-time, accurate insights can impact patient outcomes. This optimization guarantees the anonymized data remains valuable for predictive analytics, healthcare decision-making, and diagnostics. Lastly, the model’s performance is investigated with conventional methods, and the results demonstrate that this approach provides more effective solutions for this work.

3.3 Dataset details

The designed data anonymization system in IoT for the healthcare domain leverages the following dataset.

Dataset (“Healthcare Dataset”): https://www.kaggle.com/datasets/prasad22/healthcare-dataset: access date: 2025–10-21. It is a synthetic healthcare data source that is leveraged for various sectors such as data analysis, data science, and machine learning. It is designed to replicate real-world healthcare data, allowing users to practice and implement their data manipulation and evaluation skills in the healthcare sector. In this dataset, 55,500 records, including name, age, blood type, gender, date of admission, medical condition, hospital, doctor, insurance provider, room number, medication, discharge date, and so on, are available.

From this dataset, the fetched data are specified as . In addition, the total number of the fetched data elements is specified as .

An artificial dataset sourced from Kaggle was used to conduct the analysis in this research article. This dataset contains 55,500 entries. Although artificial data enables researchers to examine the efficacy of privacy-preserving techniques in a controlled setting, they may not be able to replicate real-world IoT data entirely due to its complexity.

Synthetic data generation is important because of the necessity to conduct an experiment under controlled conditions in order to measure the effectiveness of privacy-preserving mechanisms such as differential privacy. The benefit of using synthetic data is that the level of privacy can be accurately measured and guaranteed.

The chosen baselines are intended to facilitate fair comparison between the baselines and the RAOOA technique, especially on optimization efficiency, convergence properties, and utility/privacy trade-off effectiveness.

The dataset used in this research is artificial and does not fully represent the complexities of a real-world environment of cross-system hospitals. Nevertheless, the dataset allows the researchers to test the proposed privacy-preserving framework in a controlled environment.

4 Proposed data anonymization framework in the IoT healthcare domain: existing concepts

4.1 Conventional Res-BiGRU

The GRU modified the estimation strategy of the RNN’s hidden layer. It tuned the three gates in the LSTM into update and reset gates and integrated the member hidden states to control the test data’s flow at distinct times, making the technique easily retrieve the data from prior texts. However, the GRU model cannot capture the relations between the current and previous data effectively. Therefore, the BiGRU (Zhang et al., 2019) is suggested. In BiGRU, the input offers two GRUs in opposite directions, and the result is estimated by two unidirectional GRUs. The present BiGRU’s hidden layer is estimated using Equations 1–3 by the integration of the outcome of the reverse hidden layer state and the outcome of the hidden layer state forward at moments. The BiGRU is divided into two unidirectional GRUs, hence, the BiGRU’s hidden state at time is attained by weighting the reverse hidden layer state and the forward hidden state .

Here, the term indicates the nonlinear modification to the input word vector. The variables specify the weights related to the reverse hidden state and the forward hidden layer state . The variable indicates the offset value related to the hidden layer state at time .

However, the BiGRU model faces the vanishing gradient problem. To rectify this problem, the residual connections (Khan et al., 2021) are integrated with BiGRU. The Res-BiGRU includes residual connections in BiGRU to improve the training dynamics, most importantly in deep architectures. In the BiGRU, the input flows via distinct recurrent layers, which can result in issues such as vanishing gradients and performance degradation as depth increases. A residual connection resolves this by including the layer input directly into its output before passing it to the subsequent layer. This shortcut path enables gradients to flow more easily backward at training time, stabilizes learning, and supports the network retaining data from previous layers. Thus, the Res-BiGRU performs the tasks very effectively by capturing both previous and past data. Figure 3 gives the diagrammatic view of the existing Res-BiGRU model.

FIGURE 3

4.2 Conventional differential privacy

The DP (Talaei and Izadi, 2024) provides strong confirmations to secure the data in machine learning approaches. A randomized approach is considered to be differentially private if its result is effective against any variation of a single sample in the input data. The following definition defines this statement for :

: A method satisfies for two non-negative integers if for all nearby datasets such that and all subsets of range, it holds using below Equation 4

Here, the variation between two data sources , is explained as the number of records on which they vary.

It is summarized from this definition that, with a likelihood of , the outcome of a differentially private approach on two nearby data sources changes more than an attribute of . Hence, tiny values of improve the likelihood of having similar outcomes. The lower rate of the variable narrows the region for privacy protection. The smaller and are, the less the concern of privacy protection. This model considers a -random dimensional function employed on a data source, for and , a Gaussian approach with attributes that include the noise scaled to to each factor of the result, if is differentially private. The attribute defines the function’s sensitivity, explained by .

4.3 Federated learning with differential privacy

The entire client considers being included in learning. For each global iteration, the weights of the neural network designed locally are delivered to the server for collection. To guarantee for uplink attribute transmission, the Gaussian noise must be included in each individual neuron-to-neuron weight with the noise distribution , where using Equation 5

Here, the term indicates the maximum count of local attributes exposed at the uplink transmission time. The attribute for each user is the outcome of the local training task of that client . Hence, the highest sensitivity value of , explained as the maximum interval of viewing by a change in one local data source sample, is with explained as the clipping attribute , and the factor denotes the local datasets’ maximum size.

In client size, after perturbing the local weights, the server obtains the attributes and decides whether much noise must be included in the gathered attributes for downlink transmission or not. By focusing on the clients and global iterations, the Gaussian noise’s standard deviation is included in the gathered attributes to guarantee that for the downlink channel is

The Equation 6 is used for evaluating the metrics are given below to give clarity on how to measure them mathematically.

5 Proposing federated learning-based adaptive differential privacy for data anonymization in IoT healthcare

5.1 Revamped AOOA

The RAOOA is implemented in this research work.

Purpose: The RAOOA is designed for improving the performance of the AI-aided data anonymization task, specifically concentrating on optimization of noise levels in the DP model. Therefore, the information loss and the risk of re-identification are minimized, while the optimal utility is increased in the IoT healthcare data.

Reasons for choosing conventional AOBOA: The introduced RAOOA considers the features of the existing AOBOA. The conventional AOBOA (Al-Sharqi et al., 2024) is a swarm-intelligence, nature-inspired mechanism that copies the effective and hierarchical foraging character of bees in an apiary. The existing AOBOA has shown notable merits such as fast convergence, effective global search strategies, and high exploration capabilities, making it effective for complex optimization operations. Therefore, the conventional AOBOA is considered in this research.

Novelty: Although the existing AOBOA shows various strengths, it also has problems such as poor exploitation in later iterations, premature convergence, and local optima issues when handling high-dimensional data sources. To address all these problems, the RAOOA is implemented with a novel strategy. In the worker lifecycle phase, the conventional AOBOA introduces a random factor in the range of [0, 1], which can be the main cause of the problems in the algorithm because it struggles to search the solution space effectively. Therefore, by utilizing fitness values, this random factor is updated in the worker lifecycle phase. This modification improves the convergence speed, eliminates the local optima problem, and guarantees robust parameter tuning. The new formulation of the random integer is provided in Equation 7.

Here, the “current, worst, and mean fitness rates” are indicated as . In addition, the new random factor is , which is updated in the worker lifecycle phase, as shown in Equation 8.

Here, the attribute specifies the worker of the hive, and the chosen hive’s worker is indicated as . The maximum age of the worker is given as , and the age of the worker is denoted as . Furthermore, the formulated new random factor is and is derived in Equation 7.

The RAOOA is designed by eliminating the static randomness in the worker lifecycle phase of the algorithm, thus enhancing its performance in terms of scalability, stability, and optimization accuracy, making it a better choice for tuning the noise levels and balancing the privacy-utility in the sensitive healthcare IoT applications. Algorithm 1 provides the designed RAOOA’s pseudo-code.

5.2 Fitness-based factor in RAOOA

In the initial AOBOA, the updating scheme involves the use of random factors that manage the exploration phase in the optimization process. Although random factors ensure that the local optima are avoided, they may make the process unstable and may affect the optimization performance due to inconsistency in optimization behavior during iterations. In this case, the RAOOA incorporates the use of fitness-based adaptive factors in order to replace random factors. The fitness-based factors enhance the process by allowing the optimization process to be driven by solution quality rather than using random values. Better solutions affect the search more than poorer solutions, thus ensuring that the optimization process converges faster to optimal regions. On the other hand, poorer solutions have a lower effect, thus ensuring that no extra exploration is done in the process. From the optimization point of view, incorporating the fitness-based factor minimizes stochastic variability in the search process, ensuring a more controlled and more directed process.

5.3 Proposed AI-based data anonymization

Data anonymization defines the operation of masking or modifying sensitive data in datasets to eliminate the recognition of individuals while retaining the data’s usability for decision making, analysis, and predictive modeling. Its importance lies in allowing secure processing and sharing of the data, most importantly in the healthcare sector, where patient data should be protected under strict privacy regulations. Conventional anonymization approaches often face significant problems such as poor scalability, high information loss in distributed environments, such as IoT, and susceptibility to inference attacks. Most of the conventional approaches fail to maintain the trade-off between utility and privacy, degrading the accuracy and quality of the insights derived from the data.

To resolve this problem, AI-based data anonymization is utilized in this work, which addresses the problem of providing secure and utility-preserving data anonymization in IoT, especially within the sensitive healthcare domain. Initially, in the data anonymization process, the data context and privacy level prediction are performed with the support of AI.

The data context defines understanding the semantic meaning, sensitivity, and relevance of distinct attributes within the IoT data. At the same time, the privacy level prediction involves dynamically validating the degree of privacy protection needed for each data attribute based on its sensitivity and context. The conventional approaches mostly use uniform privacy policies across all data, which can result in major problems: either high anonymization, which minimizes the data utility, or inadequate anonymization, which compromises privacy. Moreover, the existing deep learning approaches, such as LSTM or GRU, tend to suffer from vanishing temporal awareness or information loss when scaled to dense models, making them ineffective for accurate context modeling in sequential IoT data. To resolve these problems, an effective AI model named the Res-BiGRU model is suggested. The data context and privacy level prediction by the Res-BiGRU model resolves the privacy concerns inherent in IoT-based healthcare devices. The Res-BiGRU model utilizes the gathered data as input.

The Res-BiGRU technique is selected in this work because of its capability to improve learning across distinct layers without affecting the quality of feature extraction. The bidirectional structure allows the technique to capture both future and past dependencies in the time sequences of IoT data, which is significant for understanding the intricate data patterns and context. In this, the residual connections support preserving the input data by adding the input directly to the output of each BiGRU layer, enabling the technique to maintain high-quality gradients and understand deeper hierarchical features without affecting performance. This architectural improvement makes the Res-BiGRU suitable for data context modeling, as it can better explain the semantic and temporal relations in the IoT data streams. Once the data context is recognized, the Res-BiGRU forecasts each attribute’s privacy level by estimating its sensitivity in relation to the nearby data context. This adaptive and dynamic privacy-level prediction is an important step beyond rule-based methods, allowing fine-grained anonymization methods. By utilizing the Res-BiGRU, the model guarantees that only the significant data parts are anonymized to a certain degree, preserving as much utility as possible for clinical decision making or downstream analytics. In healthcare IoT systems, where real-time responsiveness and data accuracy are crucial, this approach guarantees that privacy protection does not come at the cost of usability. Thus, the integration of Res-BiGRU for context and privacy level prediction offers an intelligent, scalable, and context-aware solution to data anonymization, resolving the primary complexities in privacy preservation and utility retention that conventional approaches fail to address properly. Finally, the predicted data context and privacy level are achieved.

By utilizing this AI-based data context and privacy level predicted data, the data anonymization is performed. Here, the FL-ADP is utilized, which utilizes the DP as a base model.

The DP has developed as a powerful mathematical model for privacy-preserving data sharing, in which random noise is added to the data to hide individual information. Although the DP offers strong privacy guarantees, a primary problem is the fixed noise addition, which can affect the data utility or insufficiently preserve the privacy, most importantly in dynamic IoT environments with dynamic sensitive levels.

Hence, it is significant to optimize the noise on the basis of context and the sensitivity of data attributes. The RAOOA is incorporated with the DP, which introduces an adaptive strategy. By utilizing the adaptive mechanism, the DP optimizes the noise parameters dynamically, guaranteeing optimal trade-offs between data utility and privacy preservation by eliminating problems such as premature convergence and local optima in the optimization. The objective function of the noise optimization is given in Equation 9.

Here, the optimized noise in dB is specified as , which is selected in the interval of [0–25]. Furthermore, this optimization process maximizes the optimal utility while minimizing the information loss and risk of re-identification .

To ensure consistency and completeness in evaluation, the privacy and utility metrics are formally defined as follows:

Optimal utility (OU):

where represents the utility of anonymized data and represents the utility of the original dataset.

Information loss (IL):

which quantifies the loss of useful information after anonymization.

Risk of re-identification (RR):

where is the total number of records and is the size of the equivalence class to which record belongs.

Discernibility metric (DM):

where is the number of equivalence classes and is the number of records in each class.

These formal definitions provide a standardized basis for evaluating the trade-off between data utility and privacy preservation.

Thus, the ADP, in which the noise level is tuned innovatively, is developed. However, employing ADP in the centralized setting may risk data leakage and face scalability problems. This leads to the integration of FL, a decentralized paradigm, in which the data remain on edge systems, and only encrypted technique updates are shared with the central server. FL is selected because of its capability to preserve data locality, minimize communication costs, and improve data security, which is significant for IoT healthcare systems, where the data are distributed across distinct smart systems. By integrating FL with ADP, the FL-ADP technique allows privacy-aware collaborative learning across distinct nodes without exposing the original data, while adaptively anonymizing the sensitive attributes by employing the optimized noise.

Initially, the AI-based predicted attribute contexts and required privacy levels are considered by the designed FL-ADP model, in which the RAOOA optimizes the noise levels. These anonymized outputs are employed in the FL technique, in which the distributed learning happens without centralizing the sensitive data. The combination guarantees strong privacy, low information loss, and high model effectiveness. Thus, the AI-based FL-ADP is designed for critical healthcare applications, which improves the data utility and provides context-aware anonymization that is robust against inference attacks and in compliance with data privacy regulations. Figure 4 provides the diagrammatic view of the designed AI-based data anonymization.

FIGURE 4

The proposed RAOOA-FL-ADP algorithm was implemented using the Python programming language using popular machine learning libraries such as TensorFlow/Keras for building models and NumPy for numerical calculations. Experiments have been carried out using a computer with an Intel Core i5 processor and 8 GB of RAM running under Windows OS.

5.4 Implementation details and hyperparameter settings

The clipping norm threshold value is of great importance for regulating the magnitude of update during the differential privacy process; hence, the trade-off between data utility and privacy. In this regard, the low values imply high privacy with high information loss and vice versa. Selection of hyperparameters was done on the basis of practical tuning and computational efficiency. A small population size helps reduce the computational load while ensuring enough variety in the process of search. Likewise, the number of iterations was decided in order to converge while keeping the computation time within limits. Initial testing with larger populations and more iterations revealed slightly better results but at an increased computational cost.

Fixed hyperparameters were used for baseline models, and their specific values are provided in this section to ensure reproducibility. Experiments were conducted several times, and average results were obtained to provide robustness.

In order to facilitate reproducibility and fairness in the comparisons, the hyperparameters of all baseline models have been kept the same and are provided below in Table 2. In the case of the random forest (RF) model, the number of trees is 100, the maximum depth is 20, and the minimum samples required for splitting are 2. In the case of the support vector machine (SVM) model, a radial basis function (RBF) kernel is used with C = 1.0 and gamma = 0.01. In the case of the K-nearest neighbors (KNN) model, k is 5, and the distance metric is Euclidean. For the decision tree (DT) model, the maximum depth is 20, and the minimum number of samples for each leaf is 1.

5.5 Simulation setup

Python was utilized for executing the designed data anonymization framework. The RAOOA considered 50 maximum iterations and 10 populations, and the chromosome length was determined based on the number of parameters. The conventional heuristic algorithms, such as Frilled Lizard Optimization (FLO) (Falahah et al., 2024), Skill Optimization Algorithm (SOA) (Givi and Hubalovska, 2023), Hippopotamus Optimization Algorithm (HOA) (Amiri et al., 2024), and AOBOA (Al-Sharqi et al., 2024), prediction models, such as DNN (Al-Shammari et al., 2021), 1DCNN (Mol, 2025), RNN (Morid et al., 2023) and LSTM (Liu S. et al., 2022), and the anonymization methods, such as k-Anonymity (Majeed and Hwang, 2023), L-diversity (Parameshwarappa et al., 2021), LDP (Garcelon et al., 2021), FL-DP (Talaei and Izadi, 2024), were utilized for the analysis. The designed RAOOA-FL-ADP was analyzed by distinct clipping norm threshold values: 0.15, 0.30, 0.45, 0.60, and 0.75. These threshold values control the modal update’s magnitude to balance the trade-off between data utility and privacy. The smaller norms invoke stronger privacy with higher data loss, while larger norms preserve more utility with very low privacy protection. The adaptive component is calculated on the basis of the fitness value normalization, thus ensuring that more successful solutions exert a relatively greater impact on the updating stage. This might be the limitation that influences the generalization of the results to the real healthcare system.

Some further experiments were carried out with different population numbers and iterations to study their influence on performance. It was found that an increase in these parameters beyond the chosen ones brought very small improvements in performance but greatly increased computational time.

Statistical reliability was assessed using descriptive metrics including best, worst, mean, median, and standard deviation.

Table 3 presents the statistical evaluation of the proposed model using best, worst, mean, median, and standard deviation values. Additionally, 95% confidence intervals (CI) are included to provide a more robust estimation of the variability and reliability of the model performance across multiple runs.

Metrics for predictions are used to measure the success of pattern detection by the model, whereas anonymization metrics measure privacy preservation. Such metrics are orthogonal, because the higher the prediction metric value, the higher the utility of the model. The changes made increase the transparency and clarity of experiments performed.

The metrics can be classified based on whether they belong to prediction or privacy metrics. Prediction metrics measure the accuracy of the models, whereas privacy metrics measure data privacy.

Prediction metrics measure how well the model learns the patterns, whereas privacy metrics measure the success of the anonymization process.

5.6 Performance measures

The designed data anonymization framework considers the following measures for validation.

Accuracy using below Equation 10:

Here, the true positive, true negative, false positive, and false negative rates are given as .

MPE using below Equation 11:

Here, the actual and predicted rates are given as . The number of samples is given as .

SMAPE using below Equation 12:

MASE using below Equation 13:

MAE using below Equation 14:

Root mean square error (RMSE) using below Equation 15:

MSE using below Equation 16:

NMSE using below Equation 17:

Here, the term is the actual value’s variance.

Mean absolute percentage error (MAPE) using below Equation 18:

5.7 Convergence analysis

Figure 5 presents the convergence validation of the designed RAOOA-FL-ADP compared to conventional algorithms. In this investigation, by varying the cost function rates, the efficiency is analyzed for distinct clipping norm threshold values. At the 40th iteration in Figure 5c, the cost function of the designed RAOOA-FL-ADP is minimized by 2.67% over FLO-FL-ADP, 6.25% over SOA-FL-ADP, 3.57% over HOA-FL-ADP, and 3.12% over AOBOA-FL-ADP. Hence, it is confirmed that the recommended RAOOA achieves higher convergence rates than the existing algorithms, thus guaranteeing effective noise optimization in FL-ADP. With the quick convergence rates, the designed RAOOA can explore the best solutions for the FL-ADP-based data anonymization task.

FIGURE 5

5.8 Statistical analysis

The statistical validation of the recommended RAOOA-FL-ADP is presented in Table 4. The FL-ADP is leveraged in the developed data anonymization framework, and its noise level is optimized by employing the RAOOA. Hence, its effectiveness is investigated by utilizing various statistical measures. Here, the performance of RAOOA-FL-ADP is contrasted with the conventional optimization algorithms. Also, this investigation is performed for various clipping norm threshold values. For the median factor at the 0.60th clipping norm threshold value, the RAOOA-FL-ADP’s effectiveness is improved by 2.37% over FLO-FL-ADP, 6.76% over SOA-FL-ADP, 1.74% over HOA-FL-ADP, and 2.1% over AOBOA-FL-ADP. Therefore, it is proven that the implemented RAOOA is suitable for performing noise optimization in FL-ADP compared to conventional algorithms.

For better stability of the results, multiple experiments and partitions of the data were tested for consistency of the performance results. This additional experimentation demonstrates that the suggested model is stable on different data distributions in the artificial dataset.

5.9 Performance analysis for prediction using hidden neuron counts

The Res-BiGRU technique is designed in this work for predicting the data context and privacy level. The Res-BiGRU technique is a robust model in capturing the intricate and relevant features from the input data, thus improving the predictive performance. In this analysis, the Res-BiGRU model is compared to conventional prediction techniques in Figure 6. Here, the hidden neuron counts are considered for validating the model, as it is supportive in determining the performance during training. The accuracy of the Res-BiGRU is enhanced by 5.26% over DNN, 4.21% over 1DCNN, 3.68% over RNN, and 4.73% over LSTM for the 250th hidden neuron value. Therefore, it is confirmed that the suggested Res-BiGRU is accurate and effective in predicting the data context and privacy level.

FIGURE 6

5.10 Performance analysis for prediction using activation functions

Figure 7 shows the performance investigation of Res-BiGRU-based data context and privacy level prediction compared to existing methods by considering activation functions. The investigation of Res-BiGRU with various measures helps to understand the model capabilities and also its potential in predicting the data context. For the rectified linear unit (ReLU) activation function, the designed Res-BiGRU’s MEP is minimized by 23.2% over DNN, 16.5% over 1DCNN, 19.7% over RNN, and 13.6% over LSTM. Hence, it is confirmed that the suggested Res-BiGRU achieves high predictive performance with very low error rates, supporting data anonymization in effective ways. In addition, this analysis also demonstrates that the residual connections in the Res-BiGRU minimize the gradient problems, ensuring accurate feature retrieval. The key point here is that the MAPE and MEP values presented in Figures 7c,d are remarkably close for all tested models. This issue has been checked in detail, and both measures have been calculated separately. The almost equal MAPE and MEP values (their difference equals approximately 0.01–0.03) are due to the very small magnitude and symmetric distribution of prediction errors. In particular, because MEP considers signed percentage errors, and MAPE is based on the use of absolute percentage errors, the two measures should be nearly equal if the prediction errors are low and are not biased toward a specific direction.

FIGURE 7

5.11 Overall comparative analysis for prediction

By varying distinct activation functions, the designed Res-BiGRU-based data context and privacy level prediction process is estimated and compared to existing methods. When focusing on the softmax activation function, the implemented Res-Bi-GRU’s RMSE is minimized by 90.19% over DNN, 54.9% over 1DCNN, 72.54% over RNN, and 43.13% over LSTM. Hence, it is illustrated that the recommended Res-BiGRU can predict the data context and privacy level more effectively than the previous prediction models are detailed in Table 5.

5.12 Performance analysis for anonymization

The RAOOA-FL-ADP is introduced in this research work to perform data anonymization. This technique integrates FL, DP, and noise optimization for improving the privacy and utility of the sensitive data in the healthcare domain. Though various anonymization techniques have been designed in the past years, these models face some specific problems, such as information loss. Therefore, the RAOOA-FL-ADP is designed, and its performance is investigated in Figure 8 and compared to traditional algorithms and techniques by varying the clipping norm threshold values. At the 0.75th clipping norm value, the designed RAOOA-FL-ADP’s optimal utility is increased by 69.23% over FLO-FL-ADP, 23.07% over SOA-FL-ADP, 26.15% over HOA-FL-ADP, and 7.69% over AOBOA-FL-ADP. Hence, the solutions elucidate that the introduced RAOOA-FL-ADP balances the privacy and utility trade-off without facing information loss and the risk of re-identification. Therefore, the suggested RAOOA-FL-ADP-based data anonymization is more effective than the conventional models.

FIGURE 8

In order to ensure the effectiveness of the suggested modification, the convergence process was additionally studied by comparing the proposed RAOOA and AOBOA algorithms. It can be noted that RAOOA demonstrates faster convergence and fewer oscillations.

The experiments were carried out through several independent trials in order to obtain reliable results. The confidence interval at 95% is also included to represent the significance of the results. Inclusion of the confidence intervals and the variances shows that the proposed model is stable, and there will be no possibility of overfitting the data.

5.13 Data limitations and future work

While the research presented interesting findings, its limitation is that only a single synthetic dataset was used, which may not adequately capture the complexities of actual IoT healthcare scenarios. Real-world datasets generally have elements of noise, missing data, and dynamic features that synthetic data cannot adequately reflect. The future research plan includes testing the validity of the proposed solution on real-world IoT data sets such as intelligent healthcare monitoring systems and wearable sensor networks. This will make it possible to evaluate the scalability and practicality of the model more effectively.

6 Conclusion

In this work, a new AI-based data anonymization strategy has been developed for the healthcare domain. Initially, the IoT data were fetched from a significant resource. After that, the data were provided to Res-BiGRU for forecasting the data context and privacy level, and the FL-ADP technique was applied to improve data utilization and minimize information loss. The model’s effectiveness was increased by tuning the noise in the ADP model, in which the RAOOA was utilized. Finally, the model was analyzed with the existing approaches. The information loss of the introduced RAOOA-FL-ADP was reduced by 35% over k-Anonymity, 10% over L-diversity, 95% over LDP, and 60% over FL-DP for the 0.60th clipping norm value. Hence, it has been guaranteed that the implemented data anonymization strategy can improve the data utility and eliminate information loss compared to conventional approaches. It can be concluded that the fitness-based modification in RAOOA provides better convergence stability and optimization efficiency than the conventional random-based methods. The provided implementation details and planned code release ensure reproducibility of the proposed framework.

Despite the successful results of this study and the efficiency of the introduced AI-based data anonymization framework, some limitations for further research still exist. First, it is essential to note that there is an assumption that network connectivity will be stable enough to make FL possible, and this assumption cannot be guaranteed in the real IoT environment. To solve this limitation, validation of the framework in the real-world environment is needed.

Future research will be dedicated to testing the framework in the real-world IoT environment and investigating its privacy resistance for better reliability. In addition, integration of the blockchain-based environment will be considered to provide better explainability and trust. Validation of the introduced solution based on real-world IoT datasets will be done. In addition, further investigation will include testing the proposed framework using real-world, cross-system hospital datasets. These are the changes made concerning the limitations of the dataset.

References

• 1

  Al-ShammariN. K.AlzamilA. A.AlbadarnM.AhmedS. A.SyedM. B.AlshammariA. S.et al (2021). Cardiac stroke prediction framework using hybrid optimization algorithm under DNN. Eng. Technol. and Appl. Sci. Res.11, 7436–7441. 10.48084/etasr.4277

  • CrossRef
  • Google Scholar
  • View reference in article

• 2

  Al-SharqiM.Al-ObaidiA.Al-MamoryS. (2024). Apiary organizational-based optimization algorithm: a new nature-inspired Metaheuristic algorithm. Int. J. Intelligent Eng. Syst.17 (3), 783–801. 10.22266/ijies2024.0630.61

  • CrossRef
  • Google Scholar
  • View reference in article

• 3

  AlhijaM. A.Al-BaikO.HusseinA.AbdeljaberH. (2024). Optimizing blockchain for healthcare IoT: a practical guide to navigating scalability, privacy, and efficiency trade-offs. Indonesian J. Electr. Eng. Comput. Sci.35 (3), 1773–1785. 10.11591/ijeecs.v35.i3.pp1773-1785

  • CrossRef
  • Google Scholar
  • View reference in article

• 4

  AmiriM. H.Mehrabi HashjinN.MontazeriM.MirjaliliS.KhodadadiN. (2024). Hippopotamus optimization algorithm: a novel nature-inspired optimization algorithm. Sci. Rep.14 (5032), 5032. 10.1038/s41598-024-54910-3

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 5

  AndrewJ.EuniceR. J.KarthikeyanJ. (2023). An anonymization-based privacy-preserving data collection protocol for digital health data. Front. Public Health11, 1125011. 10.3389/fpubh.2023.1125011

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 6

  ChenJ.XueJ.WangY.HuangL.BakerT.ZhouZ. (2023). Privacy-Preserving and Traceable Federated Learning for data sharing in industrial IoT applications. Expert Syst. Appl.213, 119036. 10.1016/j.eswa.2022.119036

  • CrossRef
  • Google Scholar
  • View reference in article

• 7

  ChhetriT. R.Kumar DehuryC.VargheseB.FenselA.SriramaS. N.DeLongR. J. (2024). Enabling privacy-aware interoperable and quality IoT data sharing with context. Future Gener. Comput. Syst.157, 164–179. 10.1016/j.future.2024.03.039

  • CrossRef
  • Google Scholar
  • View reference in article

• 8

  DenisN.LaurentM.ChabridonS. (2023). Integrating usage control into distributed ledger technology for internet of things privacy. IEEE Internet Things J.10 (22), 20120–20133. 10.1109/jiot.2023.3283300

  • CrossRef
  • Google Scholar
  • View reference in article

• 9

  DevineniS.GorantlaB. (2023). Security and privacy issues in internet of things (IoT) devices using COPRAS method. Intelligence2 (4), 15–22. 10.46632/jdaai/2/4/3

  • CrossRef
  • Google Scholar
  • View reference in article

• 10

  FalahahI. A.Al-BaikO.AlomariS.BektemyssovaG.GochhaitS.LeonovaI.et al (2024). Frilled Lizard Optimization: A Novel nature-inspired Metaheuristic Algorithm for Solving Optimization Problems.

  • Google Scholar
  • View reference in article

• 11

  FernándezM.JaimunkJ.ThuraisinghamB. (2022). A privacy-preserving architecture and data-sharing model for cloud-IoT applications. IEEE Trans. Dependable Secure Comput.20 (4), 3495–3507. 10.1109/tdsc.2022.3204720

  • CrossRef
  • Google Scholar
  • View reference in article

• 12

  GarcelonE.PerchetV.Pike-BurkeC.PirottaM. (2021). Local differential privacy for regret minimization in reinforcement learning. Adv. Neural Inf. Process. Syst.34, 10561–10573.

  • Google Scholar
  • View reference in article

• 13

  GiviH.HubalovskaM. (2023). Skill optimization algorithm: a new human-based metaheuristic technique. Comput. Mater. and Continua74, 179–202. 10.32604/cmc.2023.030379

  • CrossRef
  • Google Scholar
  • View reference in article

• 14

  GoelA.PujariT. (2023). Privacy-preserving IoT data aggregation in adversarial environments. Int. J. Multidiscip. Res.5 (6), 42006. 10.36948/ijfmr.2023.v05i06.42006

  • CrossRef
  • Google Scholar
  • View reference in article

• 15

  IrshadR. R.HussainS.HussainI.AhmadI.YousifA.AlwayleI. M.et al (2023). An intelligent buffalo-based secure edge-enabled computing platform for heterogeneous IoT network in smart cities. IEEE Access11, 69282–69294. 10.1109/access.2023.3288815

  • CrossRef
  • Google Scholar
  • View reference in article

• 16

  JayapradhaJ.PrakashM.AlotaibiY.KhalafO. I.AlghamdiS. A. (2022). Heap bucketization anonymity—An efficient privacy-preserving data publishing model for multiple sensitive attributes. IEEE Access10, 28773–28791. 10.1109/access.2022.3158312

  • CrossRef
  • Google Scholar
  • View reference in article

• 17

  KhanS. U.HaqI. U.KhanZ. A.KhanN.LeeM. Y.BaikS. W. (2021). Atrous convolutions and residual GRU-Based architecture for matching power demand with supply. Sensors21 (7191), 7191. 10.3390/s21217191

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 18

  KilY.-S.LeeY.-J.JeonS.-E.OhY.-S.LeeI.-G. (2024). Optimization of privacy-utility trade-off for efficient feature selection of secure internet of things. IEEE Access. 12, 142582–142591. 10.1109/ACCESS.2024.3467049

  • CrossRef
  • Google Scholar
  • View reference in article

• 19

  KumarM.MukherjeeP.VermaS.ShafiJ.WozniakM.IjazM. F.et al (2023). A smart privacy preserving framework for industrial IoT using hybrid meta-heuristic algorithm. Sci. Rep.13 (1), 5372. 10.1038/s41598-023-32098-2

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 20

  LiuH.GuT.ShojafarM.AlazabM.LiuY. (2022a). Opera: optional dimensional privacy-preserving data aggregation for smart healthcare systems. IEEE Trans. Industrial Inf.19 (1), 857–866. 10.1109/tii.2022.3192037

  • CrossRef
  • Google Scholar
  • View reference in article

• 21

  LiuS.WangX.XiangY.XuH.WangH.TangB. (2022b). Multi-channel fusion LSTM for medical event prediction using EHRs. J. Biomedical Informatics127, 104011. 10.1016/j.jbi.2022.104011

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 22

  LiuQ.HuangY.JinC.ZhouX.MaoY.CatalC.et al (2024). Privacy and integrity protection for IoT multimodal data using machine learning and blockchain. ACM Trans. Multimedia Comput. Commun. Appl.20 (6), 1–18. 10.1145/3638769

  • CrossRef
  • Google Scholar
  • View reference in article

• 23

  MajeedA.HwangS. O. (2023). Differential privacy and k-anonymity-based privacy preserving data publishing scheme with minimal loss of statistical information. IEEE Trans. Comput. Soc. Syst.11, 3753–3765. 10.1109/tcss.2023.3320141

  • CrossRef
  • Google Scholar
  • View reference in article

• 24

  MolJ. (2025). “A novel deep learning-based 1D-CNN-optimized GRU approach for heart disease prediction,”, 66. Automatika, 79–90.

  • Google Scholar
  • View reference in article

• 25

  MoridM. A.ShengO. R. L.DunbarJ. (2023). Time series prediction using deep learning methods in healthcare. ACM Trans. Manag. Inf. Syst.14, 1–29. 10.1145/3531326

  • CrossRef
  • Google Scholar
  • View reference in article

• 26

  MurakamiT.SeiY. (2023). Automatic tuning of privacy budgets in input-discriminative local differential privacy. IEEE Internet Things J.10 (18), 15990–16005. 10.1109/jiot.2023.3267082

  • CrossRef
  • Google Scholar
  • View reference in article

• 27

  NawshinF.UnalD.HammoudehM.SuganthanP. N. (2024). AI-powered malware detection with differential privacy for zero trust security in internet of things networks. Ad Hoc Netw.161, 103523. 10.1016/j.adhoc.2024.103523

  • CrossRef
  • Google Scholar
  • View reference in article

• 28

  ParameshwarappaP.ChenZ.KoruG. (2021). Anonymization of daily activity data by using ℓ-diversity privacy model. ACM Trans. Manag. Inf. Syst. (TMIS)12, 1–21. 10.1145/3456876

  • CrossRef
  • Google Scholar
  • View reference in article

• 29

  PashamS. D. (2023). Privacy-preserving data sharing in big data analytics: a distributed computing approach. Metascience1 (1), 149–184. 10.18535/mcsj/v2023.04

  • CrossRef
  • Google Scholar
  • View reference in article

• 30

  PrabhaP.ChatterjeeK. (2024). RSHealth: a ring signature scheme for identity anonymization and transaction privacy in blockchain based E-Healthcare systems. IEEE Access. 12, 117701–117720. 10.1109/ACCESS.2024.3439611

  • CrossRef
  • Google Scholar
  • View reference in article

• 31

  RenW.TongX.DuJ.WangN.LiS.MinG.et al (2024). Privacy enhancing techniques in the internet of things using data anonymisation. Inf. Systems Frontiers26 (6), 2227–2238. 10.1007/s10796-021-10116-w

  • CrossRef
  • Google Scholar
  • View reference in article

• 32

  SalimS.MoustafaN.TurnbullB.RazzakI., (2022). “Perturbation-enabled deep federated learning for preserving internet of things-based social networks,” ACM Trans. Multimedia Comput. Commun. Appl. (TOMM). 18, no. 2s, pp. 1–19. 10.1145/3537899

  • CrossRef
  • Google Scholar
  • View reference in article

• 33

  SfarA. R.NatalizioE.MazloutS.ChallalY.ChtourouZ. (2022). Privacy preservation using game theory in e-health application. J. Inf. Secur. Appl.66, 103158. 10.1016/j.jisa.2022.103158

  • CrossRef
  • Google Scholar
  • View reference in article

• 34

  SunX.YeQ.HuH.DuanJ.XueQ.WoT.et al (2023). Puts: privacy-Preserving and utility-enhancing framework for trajectory synthesization. IEEE Trans. Knowl. Data Eng.36 (1), 296–310. 10.1109/tkde.2023.3288154

  • CrossRef
  • Google Scholar
  • View reference in article

• 35

  TalaeiM.IzadiI. (2024). Adaptive Differential Privacy in Federated Learning: A Priority-based Approach. arXiv preprint.

  • Google Scholar
  • View reference in article

• 36

  WangD.RenJ.WangZ.ZhangY.ShenX. S. (2022). PrivStream: a privacy-preserving inference framework on IoT streaming data at the edge. Inf. Fusion80, 282–294. 10.1016/j.inffus.2021.11.013

  • CrossRef
  • Google Scholar
  • View reference in article

• 37

  WuJ.M.-T.TengQ.HudaS.ChenY.-C.ChenC.-M. (2023). A privacy frequent itemsets mining framework for collaboration in IoT using federated learning. ACM Trans. Sens. Netw.19 (2), 1–15. 10.1145/3532090

  • CrossRef
  • Google Scholar
  • View reference in article

• 38

  XieY.PengL.NedjahN.GuptaB. B.TaniarD.ZhangJ. (2023). Privacy protection framework for face recognition in edge-based internet of things. Clust. Comput.26 (5), 3017–3035. 10.1007/s10586-022-03808-8

  • CrossRef
  • Google Scholar
  • View reference in article

• 39

  ZhangJ.LiuF. A.XuW.YuH. (2019). Feature fusion text classification model combining CNN and BiGRU with multi-attention mechanism. Future Internet11 (237), 237. 10.3390/fi11110237

  • CrossRef
  • Google Scholar
  • View reference in article

• 40

  ZhaoS.XuS.HanS.RenS.WangY.ChenZ.et al (2023). PPMM-DA: privacy-Preserving multidimensional and multisubset data aggregation with differential privacy for fog-based smart grids. IEEE Internet Things J.11 (4), 6096–6110. 10.1109/jiot.2023.3309132

  • CrossRef
  • Google Scholar
  • View reference in article

• 41

  ZhuK.HuangL.NieJ.ZhangY.XiongZ.DaiH.-N.et al (2022). Privacy-aware double auction with time-dependent valuation for blockchain-based dynamic spectrum sharing in IoT systems. IEEE Internet Things J.10 (8), 6756–6768. 10.1109/jiot.2022.3165819

  • CrossRef
  • Google Scholar
  • View reference in article