A Survey on Secure Blockchain-Based Federated Learning for Privacy-Preserving Data Sharing and Protection Against Unauthorized Access

• This paper gives a deep exploration into blockchain-enabled FL across various domains, from health care to smart cities and agriculture, with emerging applications in autonomous vehicles and even financial systems.

• This work identifies key challenges that lie in the integration of blockchain with FL in terms of scalability, security, computational efficiency, and decentralized trust management. This work also proposes several directions for future research.

The above-mentioned objectives are covered, and various techniques of blockchain integrated FL in various applications have been discussed in the study. The arrangement of this study is as follows: Section II takes five directions about the survey and its objective that are summarized into its ideologies; Section III provides a comparison study; Section IV deals with summarizing the entire survey; Section V concludes this study; Finally, Section VI deals with a future scope.

Lakhan et al. [18] proposed a federated learning-enabled blockchain-based task scheduling (FL-BETS) framework that used different kinds of dynamic heuristics. Among these applications are health care applications with hard constraints like deadlines and soft constraints, such as resource energy consumption at the time of execution on distributed fog and cloud nodes. The FL-BETS perspective was identifying and guaranteeing privacy preservation and data fraud prevention at different levels from local fog nodes to remote clouds, all while minimizing energy consumption and delay, and achieving the health care workload deadlines. However, lack of a proper mechanism to enable the study to dynamically adjust to unforeseen changes in network topology, which arise from difficulty in reconciling dynamic network mobility with static blockchain consensus structures, affecting real-time task allocation, has to be addressed.

Baucas et al. [19] proposed a platform based on FL and private blockchain technology within a Fog Computing with the Internet of Things (fog-Iot) network. These technologies have privacy-preserving features that enable secure data within the network. The distributive nature of the fog-IoT network was used to implement an adaptive network for wearable IoT devices. A testbed was designed to test the capability of the proposed platform to preserve the integrity of a classifier. Experimental results showed that the implementation correctly maintained a patient’s privacy and integrity of a predictive service. Nevertheless, it faced a scalability problem due to large volumes of wearable devices from the inherent tension between massive IoT concurrency and blockchain’s limited throughput, leading to bottlenecks in processing and communication.

Rahman et al. [20] presented a lightweight hybrid FL framework where the blockchain smart contracts handled the edge training plan, trust management, authentication of federated nodes, and the distribution of the globally or locally trained models. The framework further managed the reputation of edge nodes and the datasets or models they upload. It supported full encryption for datasets, model training, and inference processes. The federated edge nodes performed additive encryption, while the blockchain performed multiplicative encryption to aggregate the updated model parameters. The framework was tested using a few Deep Learning (DL) applications designed for clinical trials with COVID-19 patients. However, blockchain finality times inherently lags behind FL’s rapid aggregation cycles, thus causing latency in model aggregation cycles required for efficient FL, creating a lag that disrupts training.

Alshudukhi et al. [21] designed a framework that combined FL with blockchain technology for immediate response to emergencies by utilizing historical patient data. This helped doctors make proper decisions and treat patients individually by summarizing their entire medical history to produce better treatment results. Using both technologies together ensured access to the required information while ensuring the high security of the data with good data privacy. However, the architecture didn’t address the trade-off between real-time responsiveness and computational overhead of blockchain verification well, since blockchain cryptographic validation introduces delays incompatible with emergency health care needs.

Alsamhi et al. [22] have been pioneers in the conceptual framework and technical synergy of FL and blockchain for decentralized data sharing to balance the utility and privacy of data. FL, a decentralized machine learning paradigm, enables collaborative AI model training across multiple health care institutions without the sharing of raw patient data. It forms an ecosystem of trust, security, and data integrity when integrated with blockchain. The article was able to explicate the technical underpinnings of FL and blockchain and untangle their roles in revolutionizing health care data sharing, while illustrating the possible implications that this amalgamation can have on patient care. The framework, however, faced problems when it had to handle different data formats and incompatible blockchain frameworks at the institutions, thereby restricting the system’s interoperability.

Waheed et al. [23] introduced a new hybrid approach using FL and blockchain technology for IoT-enabled health care applications with secured and privacy-preserved solutions. This approach applied the use of a public-key cryptosystem to provide semantic security in local model updates while preserving integrity through blockchain technology, as well as enforcing access control and accountability. This FL process has enabled a secure model aggregation without sharing the sensitive patient data. Such a framework has been put into practice and tested by using EMNIST datasets, and this framework could effectively preserve the privacy and security of the data while being computationally efficient. However, in such a heterogeneous IoT device capability, FL’s synchronous training does not align with the inherently diverse hardware capacities of IoT health care devices, causing imbalanced computational loads, which would degrade system performance.

Ahmed et al. [24] proposed a new method for classifying Artificial Intelligence of Medical Things (AIoMT) devices by the deployment of decentralized processing, known as “Federated Learning”. The approach included the deployment of a system on standard IoT devices and labeled Internet of Medical Things (IoMT) devices for training purposes and attribute extraction. In this manner, interconnected attributes were extracted and mapped from a global federated cum aggregation server. The methodology targeted the extraction of interdependent devices through FL with data privacy, and it strictly followed the policies of operations. Generic labels were applied to categorize devices sending medical data through normal communication channels. However, scalability was an issue in high diversity environments, due to extreme non-IID health care data that challenges stable global aggregation in FL resulting in misclassification.

Ganapathy et al. [25] presented the Block chain enabled Federated Learning for healthcare IoT (BFL-hIoT) model to ensure privacy while training deep learning models that deal with health care data in the health care IoT domain. Here, the data remained decentralized, while the global model was communicated for training by clients through a central server (CS). The proposed model aimed to strike a balance between the demand for privacy and the desire for better performance in machine learning in health care IoT. This study represented the proposed framework BFL-hIoT as a multi-class classifier that would be useful to differentiate between normal and anomalous users by labeling users through target class labels during a class classification of user data records. However, this scheme presents enormous communication overheads both to the clients and CSs, especially in low bandwidth scenarios as FL and blockchain depend on repeated, bandwidth-intensive parameter exchanges.

Ali et al. [26] suggested a blockchain-enabled FL framework along with EMRs as a novel approach that unlocks revolutionary insights into precision medicine. Blockchain-based immutability, transparency, and cryptographic protocols facilitated the conducting of FL on the distributed EMR datasets without patient privacy compromise. Integration of blockchain technology ensured data integrity, traceability, and consent management, thereby alleviating key concerns related to data privacy and security. Through the FL paradigm, machine learning models were trained collaboratively without the need for centralizing EMR data by health care institutions and research organizations. However, high computational and energy costs arose as a result of relying on blockchain, with inherently heavy cryptographic operations and FL training cycles, which are much less practical for resource-constrained environments.

Kumar et al. [27] used the integration of Permissioned Blockchain and smart contracts along with Deep Learning techniques for developing a novel data-sharing framework, named integration of Permissioned Blockchain and Deep Learning (PBDL) for data-sharing. In detail, the scheme of blockchain has been utilized in the system by registering, verifying via zero-knowledge proof, and validating the entities engaged in the communication process via consensus mechanisms. Authenticated data was then used to suggest a new DL scheme that integrated stacked sparse variational autoencoder (SSVAE) with a self-attention-based bidirectional long short-term memory (SA-BiLSTM). In this scheme, SSVAE encoded health care data into a new format, and SA-BiLSTM enhanced the attack detection process. However, the complexity of the proposed deep learning models significantly increased training and inference times, which hinder real-time application.

The Table 1 summarizes some of the frameworks that have integrated FL with blockchain for health care and IoT applications. Techniques, such as FL-BETS, hybrid FL-blockchain, and BFL-hIoT, maintain privacy, security, and decentralized data sharing while mitigating challenges such as real-time task allocation and model synchronization. Several frameworks emphasize achieving a balance between privacy and performance, for example, adaptive IoT networks and emergency response systems. Lightweight encryption and smart contracts enhance trust, authentication, and efficiency. Challenges include scalability, computational overhead, diversified data formats, and latency in model aggregation. For example, PBDL uses more advanced deep learning models but leads to increased complexity and thus hinders its real-time usage. Moreover, heterogeneous devices have also been an issue due to energy efficiency and imbalanced workloads. Most of these issues arise from inherent architectural constraints that induced synchronization delays and latency, while the protocols in blockchain introduce delays to maintain immutability and trust. These design complexities are required to be addressed to determine the limitations that continuously exist. Overall, they are very advanced frameworks but need optimization before usage.

Hai et al. [28] proposed a novel and innovative blockchain-assisted FL approach to secure data sharing in IoT Smart Cities. The process of FL was implemented in a distributed way, and blockchain enhanced the security and resilience of applications in smart cities. Through security analysis, it could be seen that the suggested approach offered comparatively better performance and remained more resistant against various security threats and vulnerabilities. The aggregation model was designed in a decentralized manner to minimize the emerging security risks of the smart city environment. However, the system was demanding significant computational resources for blockchain operations because the decentralized consensus and large-scale FL inherently scale poorly with thousands of smart-city devices, which might deter its deployment in resource-constrained IoT devices.

Abbas et al. [29] proposed an accumulative reputation-based selection (ARS) for edge-aided Unmanned aerial vehicle (UAV) networks based on a differentially private FL framework. The framework used blockchains to prevent single-point failures through the transition from centralized to decentralized control and off-chain model storage using the interplanetary file system (IPFS) with hash-keys maintained on-chain for model integrity. It reduced blockchain size due to the adoption of IPFS and local differential privacy mechanisms that prevent the leakages of privacy. Based on the ARS score, an aggregator is selected and the validators carry out the verification of model. But ARS-based aggregator selection mechanism introduced latency due to the inability of block chain to meet the rapid coordination required in UAV networks hindering the real-time operations.

Guo et al. [30] introduced a data-sharing mechanism integrating blockchain with FL for smart cities. The blockchain ensured credibility of work nodes’ performance information and developed a work node selection algorithm and a consensus incentive mechanism namely IPoQ to efficiently support FL tasks. Differential privacy technology was adopted to defend against differential attacks, and experimental results indicated the effectiveness of the proposed data-sharing mechanism. However, the IPoQ incentive mechanism is based on several predefined parameters since the blockchain consensus delays prevent real-time parameter tuning in dynamic urban environments, which lead to suboptimal consensus performance in dynamic environments of a smart city.

Qolomany et al. [31] presented a Particle Swarm Optimization (PSO)-based method to optimize hyperparameter settings of local machine learning models within the FL paradigm. The two case studies demonstrated the effectiveness of the proposed technique: one for smart city services on traffic prediction, based on an experimental transportation dataset, and the other for Industrial IoT services, with real-time telemetry dataset-based predictions for machine failures. Experiments indicated that PSO was an efficient approach in tuning the hyperparameters of deep long short-term memory (LSTM) models compared to the grid search method. However, the PSO-based optimization framework suffered from convergence problems for very complex for high-dimensional search spaces especially under decentralized constraints and, thus, gave very inconsistent results.

Kumar et al. [32] proposed a trustworthy privacy-preserving secured framework (TP2SF) for smart cities, which consist of three modules: a trustworthiness module, a two-level privacy module, and an intrusion detection module. The trustworthiness module used a blockchain-based address reputation system. The two-level privacy module made use of a blockchain-enhanced proof of work (ePoW) technique along with principal component analysis (PCA) to transform data so that inference and poisoning attacks can be prevented. The intrusion detection module used an optimized gradient tree boosting system, eXtreme Gradient Boosting (XGBoost). A blockchain-IPFS-integrated Fog-Cloud infrastructure, called CloudBlock and FogBlock, was used to deploy the TP2SF framework in smart cities. However, the two-level privacy module introduced significant processing delays during PCA transformations and ePoW computations, which are computationally intensive, thereby reducing system responsiveness in time-critical applications.

Abdel-Basset et al. [33] proposed Blockchain-based FL framework for pandemic disease diagnosis (BFLPD), which utilized the decentralized nature of blockchain technology to design collaborative intelligence for automated diagnosis without violating trustworthiness metrics, such as privacy, security, and data sharing challenges prevalent in health care systems of smart cities. The CKKS encryption was intelligently redesigned in BFLPD to ensure the secure sharing of learning updates during the training process. The proposed BFLPD provided a decentralized, secure aggregation method that protected the integrity of the global model against adversarial attacks, thus improving the overall efficiency and trustworthiness of the system. However, the approach lacked a comprehensive evaluation of scalability when handling a rapidly increasing number of participants in the decentralized network.

Muhammad et al. [34] proposed a Secure and Transparent Traffic Congestion Control System using FL to address the increasing problems of traffic congestion in smart cities. The main issues in the existing traffic congestion control systems included data privacy, security vulnerabilities, and the need for joint decision-making. FL enabled the training of models on decentralized data without compromising data privacy. Moreover, adding blockchain technology enhanced the safety, integrity, and transparency of the system. In the proposed system, FL was used for securely collecting and analyzing the local traffic data from various sources present in a smart city, without moving sensitive data out of its location. There was no adaptive mechanism for handling abrupt traffic anomalies due to unexpected events, such as accidents or emergencies, in real-time response, as FL updates and blockchain confirmations operate on slow periodic cycles.

Yin et al. [35] proposed an efficient attribute-based participant selection scheme to allow only those individuals who were eligible for the task publisher to participate in training while maintaining high privacy requirements. This improved efficiency and defended against attacks. The scheme was further extended to incentivize clients to participate in FL and to provide an audit mechanism by using a consortium blockchain. An in-depth comparison of the scheme to other schemes was presented. The evaluation results revealed that the scheme improved FL’s efficiency in making reliable participant selection possible, thus increasing the usage of FL in smart cities. However, it failed to address the biased potential due to the inherent statistical distortions caused by privacy-driven selective sampling in FL, in which the restricted participant selection criteria were introduced. This reduction in the model’s generalizability made it undesirable.

Padma et al. [36] proposed a new blockchain-based framework for the security and integrity concerns of IoT data. The framework, called SecPrivPreserve, secured its phases: initialization, registration, data protection, authentication, data access control, validation, and data sharing and download. Various mechanisms, such as One Time Passwords (OTP), encryption, and hashing, were used at different stages to enhance confidentiality, privacy, and integrity. Since the SecPrivPreserve framework was tested on a permissioned blockchain platform, tamper-proofing and non-repudiation were automatically considered. Besides, data protection used Chebyshev polynomials and interpolation. The framework was experimented with and deployed using Fabric SDK. However, the framework’s reliance on permissioned blockchain limited interoperability with other blockchain systems, as blockchains lack universal standards for identity and cross-chain communication, which could be an obstacle to its wider use in diverse IoT environments.

Far et al. [37] carried out research on the blockchain and deep reinforcement learning (DRL) integration toward optimizing mobile transmission and safe data exchange in IoT-supported smart cities. The combination of DRL and blockchain was shown to improve the performance of IoT networks by maintaining privacy and security through clustering and categorization of IoT application systems. By combining blockchain’s decentralized framework with DRL, the investigation showed how this approach addressed privacy and security issues, improved mobile transmission efficiency, and guaranteed robust, privacy-preserving IoT systems. Moreover, blockchain integration for DRL and notable applications of DRL technology were investigated. However, the approach lacks any rigorous analysis of energy consumption in resource-constrained IoT devices when deploying DRL and blockchain integration, which are inherently power-intensive processes, impacting system efficiency.

The Table 2 depicts some blockchain-FL methods that have been used to tackle privacy, security, and efficiency in smart cities and IoT environments. Decentralized control, IPFS for model integrity, differential privacy, and encryption were some of the techniques used to enhance system resilience against threats. Optimization methods like PSO improved model performance, while frameworks like TP2SF and SecPrivPreserve ensured privacy and intrusion detection. Applications included traffic congestion control, automated diagnosis, and efficient participant selection, which were showing improved data sharing and decision making. However, the problems of high computational demands, latency in real-time operations, scalability issues, and energy inefficiency in resource-constrained devices remain significant obstacles to wide-scale adoption and efficiency. These challenges arise from the properties of the system, where the blockchain integration induces latency in most decentralized smart-city deployments, while ensuring security. Similarly, FL scalability issues occur due to heterogeneous devices, non-IID data, and frequent communication iterations, which are intrinsic to real-world IoT environments.

Shen et al. [38] presented a Blockchain-Assisted Federated Learning (BAFL) driven Support vector Machine (SVM) framework to accomplish efficient data sharing and privacy protection. BAFL-SVM was composed of FedSVM-RiceCare module and FedPrivChain module. Specifically, the model in FedSVM-RiceCare was trained by using FL and SVM to enhance experimental accuracy. In FedPrivChain, homomorphic encryption and a secret-sharing scheme were used to encrypt the local model parameters before uploading. Extensive experiments on a real-world dataset of rice pests and diseases were conducted, and the experimental results showed that the framework not only ensured secure data sharing but also achieved higher recognition accuracy compared to other schemes. But this structure needed much computation to carry out encryption and decryption, due to the requirement of significant processing, which make them unsuitable for low-power or resource-constrained devices.

Praharaj et al. [39] presented a multi-layered architecture with DT to emphasize the security challenges in cooperative smart farming. They further proposed a hierarchical federated transfer learning framework to handle and mitigate security threats against collaborative smart farming. In this approach, FL-based anomaly detection (AD) was used, running on edge servers, and allowed AD models to run locally without exposing farm data. This localization had outstanding generalization ability and greatly increased the detection of unknown cyber-attacks. A hierarchical FL structure was used, where aggregation support at various levels was supported to foster multi-party collaboration. Furthermore, the approach had the essence of Convolutional Neural Networks (CNNs) and LSTM models, which was further enhanced through transfer learning. However, the hierarchical structure caused much latency and communication overhead, while multi-level aggregation introduced unavoidable synchronization delays across layers, which negatively impacted anomaly detection in a real-time application.

Singh et al. [40] suggested a new architecture using the combination of blockchain technology and FL in enhancing the security of smart farming systems (SMSs) against external threats. Integrating blockchain technology provides authenticity of data and maintains data storage transparently, and the FL facilitates collaborative training models without compromising data privacy. The architecture utilizes an ensemble layer (EL) to train data from each smart land while maintaining safety along with privacy measures. Each device used FL techniques on smart land to collaboratively train intrusion detection models where the confidentiality of sensitive information was maintained. Data was aggregated at the Authentication Layer, and the PoAh Consensus Algorithm was utilized for authenticating data on smart land. However, the use of consensus algorithms led to scalability problems, whose complexity grows with the network size, especially in scenarios with a large number of participating nodes.

Vimalajeewa et al. [41] proposed a novel joint FL model based on Neural Network (NN) and Partial Least Square (PLS) regression (FL-NNPLS). Its predictive performance was evaluated under sequential and parallel updating-based FL algorithms in the context of smart farming for milk quality analysis. Smart farming is a fast-growing industrial sector requiring effective analytics platforms to enable sustainable farming practices. The FL-NNPLS approach outperformed the centralized approach and showed state-of-the-art performance. Although the NN-based models have various advantages, their use in FL settings was limited by thin clients with low computational capabilities and high-dimensional data with numerous model parameters. However, the model faced difficulties in handling dynamically changing data distributions, which led to inconsistencies in prediction accuracy over time because of distributed clients with heterogeneous capabilities and local data distributions, making consistent model updates fundamentally difficult in decentralized settings.

Leduc et al. [42] focused on the design, development, and application of innovative methods to use modern information and communication technologies, such as the IoT and machine learning, to move toward more sustainable agricultural and farming practices. This paper highlighted that most blockchain-based farming frameworks concentrated on food tracking and traceability. There is scant research done to focus on the design of digital markets to enable trade among farmers and other third-party participants or for conducting performance reviews of the proposed frameworks. However, dependency upon stability and high speed in a network has hampered connectivity in most areas related to rural farming and is mostly not applicable there due to the lack of network reliability as in blockchain and FL.

Hasan et al. [43] proposed a blockchain-based solution that harnessed intrinsic features of this innovative and disruptive technology to tap into transparency, reliability, authenticity, and accountability within an agricultural supply chain. Furthermore, the solution removed misleading food labels on fresh produce by providing consumers with the ability to trace and verify the origin and authenticity of the source of food transparently. The sustainable solutions implemented were clear, and all the standards met were from different sustainable food certifications and agencies, including organic, non-Genetically Modified Organisms (GMO), and Good Agricultural Practices (GAP). The solution failed to scale to larger agricultural ecosystems, which required maintaining transparency and traceability across multiple stakeholders, hence posing integration problems with complex and diverse farming networks.

Chaganti et al. [44] suggested a cloud-enabled smart-farm security monitoring framework for monitoring the status of devices and sensor anomalies effectively to counter security attacks using behavioral patterns. Further, a blockchain-based smart-contract application was implemented to store the security-anomaly information safely and proactively counter the same attacks on other farms in the community. The security-monitoring-framework prototype for smart farms using Arduino Sensor Kit, ESP32, Amazon Web Services (AWS) cloud, and smart contract on the Ethereum Rinkeby test network monitored and responded to latency within the network. Its disadvantage was that it completely depends on a single service in the cloud, which linked the system with provider’s availability and performance. This caused an increased risk of being prone to service outages as well as vendor lock-ins with reduced system robustness.

Kumar et al. [45] presented a secure privacy-preserving framework (SP2F) for smart agricultural UAVs. The SP2F framework had two major engines: a two-level privacy engine and a deep learning-based anomaly detection engine. The two-level privacy engine in the framework comprised a blockchain and smart contract-based ePoW designed for data authentication to prevent data poisoning attacks. A Sparse AutoEncoder or SAE was applied to data, transforming it into a new form, encoded, to hinder inference attacks. Here in the anomaly detection engine, a stacked long short-term memory (SLSTM) was trained on the results of the offered two-level privacy engine, applying them on two publicly available datasets based on IoT, namely, ToN-IoT, and IoT Botnet. However, the energy and computational limits of UAVs make real-time encryption, anomaly detection, and SLSTM processing challenged the framework in handling UAV hardware constraints, which might lead to reduced efficiency in resource-constrained scenarios.

Pranto et al. [46] identified all the various features and opportunities of using blockchain along with smart contracts with the IoT-enabled integration in the pre-harvesting and post-harvesting parts of the agriculture segment. They have proposed a system, based on blockchain as its main infrastructure, with the involvement of IoT devices at field-level data collection and further using smart contracts for regulations across all contributing parties in an interaction. The implementation diagrams for this system were shown with a step-by-step explanation. Costs of gas for all operations were also added in order to provide a clearer vision of costs. However, in multi-party disputes, blockchain’s immutability and deterministic smart contracts limit flexibility in resolving nuanced conflicts and, therefore, has proved less effective in resolving complex contractual disputes in agricultural cases.

Sakthi and DafniRose [47] evaluated the performance of a blockchain-assisted secured smart agriculture system by recording the execution time and communication overhead. Private blockchain services were developed to authenticate users and other devices in the vehicular network, and experiments have been conducted. Blockchain was able to provide authenticity, security, and privacy for the users and other actors in the IoV. Different days in the cloud environment recorded various mobile user query response times. TID was the abbreviation of Transaction Identification Number and denoted the time taken by the cloud service to process the query and send back a response to the internet user. However, central cloud environments raised dependency-related issues of single points of failure and potential performance bottlenecks. The dependency on a central cloud for blockchain authentication is a structural issue.

Table 3 presents exciting new applications of blockchain, FL, IoT, and advanced algorithms to smart agriculture. Techniques used, such as FL-SVM and hierarchical federated transfer learning, improved data security accuracy and anomaly detection, but suffer from high computational demands as well as latency. However, with the integration of blockchain with FL and IoT, it had issues with scalability and resolution of multi-party conflict; with advanced models such as FL-NNPLS and SP2F, the state-of-the-art anomaly detection and data protection came into existence, with limitations in handling dynamic data and resource-constrained environments. The studies notwithstanding, the contributions underscore dependence on stable networks and centralized systems, which significantly increases the risks for rural or large-scale applications. Most of these limitations are due to inherent technological principles, since, the cryptographic operations of blockchain and iterative updates of FL are naturally resource-intensive, which increases the computational costs. Additionally, integration of decentralized learning in rural, bandwidth-limited agricultural settings induces consensus delays and communication overhead.

Ashraf et al. [48] proposed a new FL framework, FedDP, to address energy theft detection. This framework allowed different clients to experience on-device prediction with reduced communication overhead while learning from the experiences of other clients through a CS. In addition to the above, a federated voting classifier (FVC) was proposed using novel majority voting-based consensus by traditional machine learning classifiers, including RFs, k-nearest neighbors, and bagging classifiers for the accurate detection of energy theft. Based on their knowledge, these classical ML classifiers had never been applied in a federated way for the detection of energy theft in SGs. However, the design was scalable and had scalability issues while integrating a large number of heterogeneous clients with varied computational resources, as FL inherently requires synchronization across devices with different capabilities.

Zhang et al. [49] introduced a privacy protection scheme for smart meters in decentralized smart home environments based on a consortium blockchain. The scheme mitigated the security issues that included single point failure and malicious data tampering with the use of elliptic curve point multiplication in a certificateless aggregated ring signcryption algorithm, which was reduced to computing and communication costs. A consortium blockchain was used for the distributed storage of user data to eliminate issues of single-point failure and data tampering. Theoretical analysis showed that the scheme ensures user privacy with confidentiality and unforgeability, whereas experiments reveal decreased computing and communication costs. However, the method was not found to work effectively in achieving interoperability among various blockchain frameworks and legacy smart grid infrastructures due to its heterogeneous nature, which is difficult to standardize.

Yilmaz et al. [50] handled the privacy violation of energy usage data acquired from the smart meters. They come up with a new novel solution to support the prevention of privacy along with supporting the analytics on energy data. They experimented the implementation of occupancy detection attacks, using a deep neural network, with promising high-accuracy results. The introduced framework is termed as adversarial machine learning occupancy detection avoidance-B (AMLODA-B). This framework employed a LSTM model within the smart metering infrastructure to prevent leakage of personal information. However, the reliance on deep learning models increased the computational load, which was intrinsic to LSTM-based privacy preservation on resource-constrained smart meters.

Fan et al. [51] presented a decentralized privacy-preserving data aggregation scheme for the smart grid through blockchain. They applied the leader election algorithm for the selection of the best mining node in the residential zone, which acts as a mining node that constructs the block. A node, within this framework, employed an algorithm from the Paillier cryptosystem for power consumption data aggregation; they used the Boneh-Lynn-Shacham (BLS) short signature along with SHA 256 function for privacy in the aggregation process and assurance of data integrity, respectively. This scheme protected the privacy of users while achieving decentralization, eliminating the requirement for Trusted Third Party (TTP) or Certificate Authority (CA). Security analysis showed that the scheme satisfied the privacy and security requirements for smart grid data aggregation. However, the dependency on leader election algorithms led to uneven load distribution and increased latency under heavy network traffic conditions, which was an inherent limitation of decentralized mining-based aggregation schemes.

Lu et al. [52] suggested a light weight Edge blockchain-assisted Data Aggregation scheme (EBDA) for preserving privacy for the smart grid. The architecture applied edge computing and blockchain across a three-layered model that supported a two-tier data aggregation model with a significant increase in efficiency as well as security. In theoretical analysis and simulation, they showed that EBDA has strong resistance to network attacks with lower computation cost and also reduces communication overhead than earlier schemes. However, with edge blockchain adoption, synchronization became a problem between the different edge nodes, especially in high-frequency updates, because coordinating frequent updates across distributed nodes inherently introduces delays.

Jithish et al. [53] proposed an FL-based anomaly detection scheme for smart grid in which ML models have been trained locally on a smart meter without sharing with a CS, keeping the user’s privacy into consideration. In the new approach, a global model was downloaded from the server at the smart meters for on-device training. After local training, the local model parameters are sent to the server with the intention of improving the global model. Model parameter updates were secured from the adversaries using the Secure Sockets Layer/Transport Layer Security (SSL/TLS) protocol. Using the standard datasets, they experimentally probed the performance of anomaly detection in FL and found that FL models could attain the same level of anomaly detection performance as centralized ML models but without compromising users’ privacy. However, this reliance on SSL/TLS did not fully address adversarial poisoning of parameter updates during aggregation, because securing parameter transmission alone failed to address malicious updates in FL, which was a fundamental limitation of current aggregation methods.

Su et al. [54] presented a secure FL-enabled Artificial Intelligence of Things (AIoT) scheme for private energy data sharing in smart grids with edge-cloud collaboration. Specifically, the authors proposed an edge-cloud-assisted FL framework for communication-efficient and privacy-preserving energy data sharing of users in smart grids. Based on the consideration of non-IID effects, they presented a local data evaluation mechanism in FL and formulated two optimization problems for Energy Data Owners (EDOs) and ESPs. Because the knowledge of multi-dimensional user private information is very scarce in practical scenarios, they proposed a two-layer deep reinforcement learning-based incentive algorithm to encourage EDOs’ participation and high-quality model contribution. However, this incentive mechanism faced challenges from a scalability perspective since incentive mechanisms struggle to handle a tremendous number of users and energy providers, which significantly caused a problem from the computation aspect.

Fekri et al. [55] suggested FL for load forecasting using smart meter data. This approach allowed training a single model on all the participating smart meters without sharing local data. Two alternative FL approaches were tested: FedSGD, which executed one step of gradient descent on clients before aggregating updates on the server, and FedAVG, which executed multiple steps before aggregating. Specifically, residential consumers are heterogeneous, and therefore, it is difficult to train one model because the load profiles were different for each consumer. The results indicate that Federated Averaging (FedAVG) performed better in terms of accuracy compared to Federated Stochastic Gradient Descent (FedSGD), which also required fewer communication rounds. However, diversity in the consumption patterns of residential often resulted in suboptimal convergence for households whose consumption patterns were unique or extreme, which made a single federated model unable to fit all clients accurately.

Li et al. [56] proposed a false data injection attack (FDIA) detection method based on secure federated deep learning by combining the transformer, FL, and the Paillier cryptosystem. The transformer, as a detector deployed in edge nodes, delved deep into the connection between individual electrical quantities using its multi-head self-attention mechanism. This approach utilized the data of all nodes in a FL framework to collaboratively train a detection model by keeping the data local during training and preserving data privacy. To improve the security of FL, an innovative scheme for secure FL was designed with the combination of the Paillier cryptosystem and FL. However, the computational overhead of the Paillier cryptosystem increased real-time anomaly detection latency much more than expected.

Shrestha et al. [57] proposed an anomaly detection system to detect threats and utilized FL to address issues of data silos and data privacy. The framework identified anomalies in industrial data gathered from remote terminal devices deployed at substations in the smart electric grid system. The anomaly detection system involved LSTM, autoencoders, as well as mean standard deviation and median absolute deviation approaches to identifying anomalies. To improve the security and privacy properties of the proposed framework, they implemented homomorphic encryption along the lines of the Paillier algorithm. However, the use of homomorphic encryption added considerable computational and energy costs, as the encryption itself is inherently resource-intensive in privacy-preserving anomaly detection, which were impractical for edge devices with limited resources.

Table 4 summarizes several privacy-preserving and anomaly detection techniques for smart grids. Many of these research studies have been conducted focusing on FL, consortium blockchain, and edge computing to address the above issues such as energy theft detection, privacy protection, and false data injection attacks. FL provides a decentralized solution, allowing local model training to protect user privacy. Scalability and computational load are big concerns, especially for devices with limited resources. Data tampering on blockchain and elliptic curve-based cryptosystems mitigates, but the difficulty remains in interoperability with legacy systems. Another, deep learning methods of LSTM and autoencoders suffer from computational load; these cannot be used directly at the edge device level due to computational intensity. In summary, the techniques are enhancing security, privacy, and efficiency, but require further optimization in terms of scalability and computational demands. These drawbacks are attributed to fundamental system constraints, such that decentralized architectures induce consensus latency, cryptographic overhead, and leader election delays. Similarly, FL struggles with heterogeneous clients and high-frequency data in the distributed learning systems design.

Zhao et al. [58] proposed a FL system, which incorporated a reputation mechanism to aid home appliance manufacturers in machine learning model training based on customers’ data. The use of this system enabled the manufacturers to predict customers’ requirements and consumption behaviors. The workflow of the system consisted of two stages. In the first stage, customers used mobile phones and Mobile Edge Computing (MEC) servers to train an initial model offered by the manufacturer. The clients started to collect data of these domestic appliances from their phones; they then downloaded the initial models, which then gave them the required training locally on these sets. Once they obtained their local models, customers started to digitally sign those models and uploaded them on to blockchain. The contribution in global model by those was uniform for all diverse customer quality which caused its possible bias in global models, because FL inherently assumes equal influence, ignoring client data quality heterogeneity.

ur Rehman et al. [59] proposed a new idea of fine-grained FL to decentralize shared ML models on edge servers. They formally extended the definition of the fine-grained FL process in mobile edge computing systems and defined the core requirements of fine-grained FL systems, which include personalization, decentralization, incentive mechanisms, trust, activity monitoring, heterogeneity, context-awareness, model synchronization, and communication and bandwidth efficiency. In addition, they proposed the concept of blockchain-based reputation-aware fine-grained FL to ensure trustworthy collaborative training in mobile edge computing systems. However, the fine-grained FL approach faced problems with managing high computational overhead during model synchronization, especially in heterogeneous edge environments with diverse client capabilities.

Li et al. [60] presented a decentralized FL framework, based on blockchain and known as the Blockchain-based federated learning framework with committee consensus (BFLC). The framework exploited blockchain for global model storage and local model update exchange. To facilitate BFLC, an innovative committee consensus mechanism was developed, which brought to the fore diminished consensus computation, resulting in reduced malicious attacks. The scalability of BFLC, including theoretical security, storage optimization, and incentives were also discussed. However, the committee-based consensus mechanism was prone to delay model updates for time-sensitive tasks under heavy network traffic, as distributed consensus is inherently slower than centralized aggregation.

Zhang et al. [61] designed a blockchain-based smart agent system architecture and applied it in FL. The FL technique trained models across multiple devices or servers holding private data samples without exchanging the data. Locally the trained results were aggregated by a centralized server in a privacy-preserving way. The proposed smart agent model created a new architecture composed of agents forming a blockchain network. A fully decentralized, privacy-preserving smart agent blockchain-FL framework was proposed where every smart agent acted as a peer of a blockchain network and participant of an FL task. Finally, artificial neural network demonstration was carried out to validate the framework’s efficacy. With each increase in the number of smart agents, the complexity of the architecture increased, thereby causing problems in managing and synchronizing the decentralized network, because a fully decentralized FL inherently required coordination among all peers.

Wang et al. [62] designed an autonomous blockchain-empowered privacy-preserving FL framework. Here, mobile edge computing technology was implemented in the Internet of Vehicles (IoV) system. Techniques related to differential privacy were utilized to manage privacy. To ensure honesty of the participants, a malicious updates removal algorithm using a self-reliability filter was designed. Concurrently, the authors introduced a double aggregation frame to reduce communication overhead while maintaining the quality of the training model. In blockchain, an FL participant management system was developed such that any models uploaded there cannot be changed due to the recording of quality. Nonetheless, the differential privacy techniques have reduced model performance, leading to a trade-off between preservation of privacy and training accuracy.

Qi et al. [63] proposed a consortium blockchain-based FL framework to enable decentralized, reliable, and secure FL without a centralized model coordinator. In the proposed framework, the model updates from the distributed vehicles were verified by miners to prevent unreliable model updates and were then stored on the blockchain. Besides, to further protect model privacy on the blockchain, a differential privacy method with a noise-adding mechanism was applied to the blockchain-based FL framework. Numerical results illustrated that the proposed schemes could effectively prevent data poisoning attacks and improve the privacy protection of model updates for secure and privacy-preserving traffic flow prediction. However, it is noteworthy that relying on miners for verifications can lead to delaying model update processing and impair real-time prediction performance, especially in critical traffic applications, as blockchain consensus and differential privacy inherently introduce latency in time-sensitive traffic applications.

Wang et al. [64] proposed a blockchain-based privacy-preserving FL scheme named BPFL that utilizes blockchain as the underlying distributed framework for FL. The Multi-Krum technology was enhanced, and homomorphic encryption was integrated to achieve ciphertext-level model aggregation and model filtering in order to allow the verification of the local models and still achieve privacy preservation. Moreover, they designed a reputation-based incentive mechanism to guide users in the IoV to actively participate in FL and practice honesty. Through security analysis and performance evaluation, the results indicated that the presented scheme could meet the proposed security requirements and improve the FL model performance. However, the introduction of homomorphic encryption will probably add significant computational overhead that impacts the efficiency of the system in low-resource computational environments in privacy-preserving FL.

Moulahi et al. [65] applied classification approaches to Vehicular Ad Hoc Network (VANET) and intelligent transport systems (ITS) cyber-threat detection at the vehicle level. Then, they used blockchain and applied an aggregation strategy to different models. The models of the previous step were uploaded into a smart contract, and the updated models were returned to the vehicles. They also conducted an experimental study to measure the effectiveness of the proposed prototype. In this paper, the VeReMi dataset was divided into five parts in a balanced manner for the experimental study. Classification techniques were executed by each vehicle separately, and models were generated. After the aggregation of the models in the blockchain, they were returned to the vehicles. However, model aggregation in the blockchain might be slow as the distributed ledger consensus is inherently time-consuming, which would be a counter to the fast response required in a high-speed vehicle network.

Hua et al. [66] proposed a method on blockchain-based FL to implement asynchronous collaborative machine learning between the distributed agents that owned data. This method performed distributed machine learning without a trusted CS. The blockchain smart contract was used to manage the whole FL process. The FL approach employed a support vector machine-based intelligent control model based on historical driving data retrieved from real heavy haul rail systems. A mixing kernel function that was constructed of polynomial and radial basis functions (RBFs), with the help of a dynamic weight factor varying with train speeds improved model accuracy. However, since SVM is employed, its dependency may have limitations toward scalability in larger complex systems that exist within real rail systems, because traditional kernel-based models do not scale efficiently with large datasets or many agents.

Durga et al. [67] presented a novel framework based on blockchain and the FL model. The FL model addressed reduced complexity, while blockchain helped in maintaining privacy in distributed data. More precisely, the proposed FL ensembled deep learning blockchain model framework collected data from different medical health care centers and developed the model with a hybrid capsule learning network. It made predictions accurately with privacy preservation and sharing, but the complexity of ensemble deep learning models may lead to increased computational requirements and performance issues in resource-constrained health care environments with limited processing power.

Several FL systems, leveraging blockchain, are presented in Table 5, for various applications, ranging from home appliance prediction and medical health care to vehicular networks. These systems exploit decentralized models with privacy-preserving techniques, such as differential privacy and homomorphic encryption. The blockchain further ensures secure updating of models and participant verification, whereas in some systems, reputation mechanisms promote honest collaboration. However, several challenges are associated with FL, such as high computational overhead during model synchronization, mainly with fine-grained FL and SVM-based models, slow aggregation due to network delays, and trade-offs between privacy and model accuracy. Some of the frameworks face scalability problems, such as complexity in synchronization and computational requirements, especially in resource-constrained environments. These issues persist from fundamental characteristics of blockchain and FL rather than due to implementation flaws, which introduced computational and communication delays. Although these systems have limitations, they have the potential to significantly improve model accuracy, privacy, and real-time performance in decentralized settings. Future work needs to address these challenges for the broader deployment. As the research works are tested using software, the simulation results are utilized for comparative analysis, which is explained below.