Blockchained Federated Learning for Internet of Things: A Comprehensive Survey

FL is a distributed ML framework [80] involving \(N\) training participants and an aggregator. Participants, such as mobile devices, utilize their local datasets \(\mathcal {D}\) for the training process and share their model parameters instead of their raw data. Meanwhile, the aggregator, such as a server, aggregates the shared local models as a global model. The central aggregator acts as a model coordinator rather than a data repository for the local data of the participants, preserving data privacy [152]. The structure allows the global model to benefit from diverse data sources without learning individual datasets, as raw data is never shared.

In a typical FL process, there are four steps involved. By using \(\lbrace P_1, \ldots , P_i, \ldots , P_N\rbrace\) to denote the \(N\) training participants, the typical FL process is shown in the Figure 3 and divided into the following parts:

Then, the updated model is returned to the participants, and the next round of training begins. These processes continue to loop until the model reaches the expected performance.

The model update in each loop is determined by the choice of the aggregation algorithm used in the FL process. One of the most commonly used aggregation algorithms is FedAvg [80], which performs aggregation by computing the average during the FL process. Specifically, FedAvg calculates the shared parameters \(w_G\) as follows:where \(\vert D_i \vert\) represents the number of local training data in the dataset \(D_i\) of participant \(P_i\).

The FedAvg algorithm has limitations, such as the need to synchronize all updated parameters at each iteration and the consideration of dataset size in weight calculation. To address the limitations, several variants of FedAvg have been proposed to improve the effectiveness of aggregation. Reisizadeh et al. [103] introduce FedPAQ, which allows for multiple local updates before sharing parameters and controls participant selection. Li et al. [61] develop the FedProx algorithm, which uses a proximal term to reduce the computing consumption of heterogeneous data. Wang et al. [129] improve the FedMA algorithm, which applies a Bayesian non-parametric mechanism to adjust the model size based on distribution heterogeneity.

As the underlying support of Bitcoin, Blockchain is a distributed ledger technology that uses cryptographic techniques to secure and maintain a decentralized database [48]. Blockchain is designed to provide independent internal verification, communication, transmission, and storage while maintaining a reliable and transparent environment [156]. This technique has the potential to meet various data requirements as it allows any peer to add new data and maintain synchronized information according to specific rules.

The Blockchain is structured as a series of blocks that store transactional information. Each block is comprised of two parts, i.e., the header and the body, as shown in Figure 4. The block header includes hash values of the previous block and its own, enabling the blocks to link and form a continuous chain [111]. The Merkle Root Hash locks all the transactions in the block such that the transactions cannot be tampered without changing the root hash. The Nonce field reflects the Blockchain consensus works. The body of the Blockchain holds detailed information about transactions, which are cryptographically secured, ensuring the data it contains is immutable and tamper-resistant once a block is added to the chain [123].

The features of Blockchain [81] have led to rapid development in existing industries, which can be described as follows:

Blockchains have already demonstrated their usefulness in the context of IoT [74, 75, 135], and the capacity of Blockchains has been analyzed in [132, 133, 153] for IoT applications. According to the application scenario, the Blockchain can be classified into three types [6], as follows:

A detailed comparison between the three Blockchain types is shown in Table 1. In terms of the consensus process, all nodes of a public Blockchain can participate, while the consensus of a private Blockchain is controlled by a single organization. A consortium Blockchain expands on the private Blockchain to include multiple organizations. Correspondingly, a public Blockchain has complete decentralization, and its access is public without requiring permission. By contrast, private and consortium Blockchains, on the other hand, are only partially decentralized and more controlled, where nodes need permission to access.

The BlockFL model combines FL and Blockchain technologies, which can offer innovative solutions in various sectors [42]. By monitoring, recording, certifying, and coordinating the FL process, BlockFL offers the following advantages:

FL also enhances Blockchain by contributing crucially to its consensus algorithms. The models and training data shared through FL, as detailed in sources [17, 136], improve the incentive structures of Blockchain. Additionally, the way of securing raw training data in FL increases the data privacy aspect of Blockchain technology.

Architecture. A BlockFL system model [50] consists of two parts: the local learning process (running on mobile devices) and the integrated calculation process (implemented on the Blockchain). In the BlockFL system, there are two main actors: participants (i.e., mobile devices) who use their local datasets to learn preliminary models, and miners in the Blockchain who verify models and facilitate aggregate calculations. The participants and miners can be either the same or different entities. The process of a BlockFL model is shown in Figure 5. Compared to traditional FL, BlockFL introduces a more complex process by adding miner validation and leader election, leveraging Blockchain technology to replace the role of traditional aggregators. The uploaded and downloaded global models in BlockFL are stored in secure blocks, and model aggregations are completed through miner campaigns. This eliminates the dependence on unreliable aggregators in FL, reducing associated risks and improving the security and trustworthiness of the overall process. The BlockFL system can be described as follows:

Why BlockFL. BlockFL is a pioneering approach that integrates the privacy-preserving structure of FL with the secure and transparent ledger system of Blockchain. While FL decentralizes model training across various nodes, ensuring data privacy by keeping it local, it typically relies on a central server for model aggregation, which can be a bottleneck [65]. Blockchain, known for its immutable and auditable transaction records, provides a secure and decentralized system but can be limited by scalability and energy consumption [25]. BlockFL capitalizes on the strengths of both, utilizing the decentralized ledger of Blockchain to enhance the security and trustworthiness of decentralized model training in FL. As shown in Table 2, we observe distinct tradeoffs across efficiency, storage, communication, and computational overhead in FL, Blockchain, and BlockFL. FL offers efficient local computations but faces limitations during global aggregation, while the efficiency of Blockchain is dependent on post-mining processes. BlockFL achieves a balance by integrating both the efficiency of FL and the overhead of Blockchain.

The storage, communication, and computation costs are evaluated in Table 2 following the BlockFL architecture, in which each local model is certified via a single Blockchain transaction. This could be done by saving the local model in the InterPlanetary File System (IPFS) and then embedding the hash and reference of the local model in a transaction. We also consider a synchronized FL process with \(R\) iteration. In each interaction, every FL participant contributes a local model by training on the local dataset, and then the local models are aggregated. The storage costs of FL, Blockchain and BlockFL are \(O(N_FD+N_FM)\), \(O(N_BTS)\), and \(O(N_FD+N_FRM+N_BTN_FR)\), respectively, where \(N_F\) is the number of FL participants, \(D\) is the average size of a local dataset, \(M\) is the size of an FL model, \(N_B\) is the number of Blockchain peers, \(T\) is the average transaction size, \(S\) is the number of transactions on the Blockchain. In the case of BlockFL, all \(N_FR\) local models are logged, i.e., \(S=N_FR\), leading to \(O(N_FRM)\) model storage and \(O(N_BTN_FR)\) Blockchain transaction storage costs across Blockchain peers. The communication costs of FL, Blockchain and BlockFL are \(O(N_FMR)\), \(O(N_BTS+f(N_B)S)\), and \(O(N_BN_FRM+N_BTN_FR+f(N_B)N_FR)\), respectively, where \(f(N_B)\) is the consensus communication cost of a transaction among \(N_B\) Blockchain peers (which varies according to the consensus protocol). In FL, \(N_FR\) local models are transferred to the aggregator. In BlockFL, \(N_B\) Blockchain peers need to learn \(N_FR\) models and model transactions and then run consensus on the transactions. The computation costs of FL, Blockchain and BlockFL are \(O(N_FRC_L+RC_A)\), \(O(N_BS+g(N_B)S)\), and \(O(N_FRC_L+N_BRC_A+N_BN_FR+g(N_B)N_FR)\), respectively, where \(C_L\) is the local learning cost, \(C_A\) is the aggregation costs, \(g(N_B)\) is the consensus computation cost of a transaction among \(N_B\) Blockchain peers depending on the consensus protocol. In BlockFL, the local training process is similar to the local training in FL, but the aggregation is independently executed by \(N_B\) peers with an aggregation cost of \(O(N_BRC_A)\). Blockchain peers also need to verify and run consensus on \(N_FR\) transactions.

Technical Features. Table 3 provides a comprehensive summary of the technical features of various BlockFL models, analyzing and comparing them based on factors such as training synchronization, chain structure, consensus mechanisms, and permission. This analysis offers valuable insights into the current state of development of popular BlockFL models.

However, the integration of Blockchain and FL introduces new security and privacy challenges. Unlike traditional FL, where local models are only shared between participants and the central aggregator, BlockFL allows all participants to access local models from each other, essentially elevating the knowledge of the participants to the level of the central aggregator. Consequently, FL participants can launch attacks that were previously exclusive to central aggregators in traditional FL, such as attacks based on consensus algorithms [35] and inference attacks [115], where attackers infer training data from model updates. The open network also raises concerns about Intellectual Property (IP) protection. Participants can learn from the local models of others and then manipulate theirs to falsely claim learning contributions. Moreover, Byzantine failures, a common threat in distributed systems, persist in BlockFL systems. For example, malicious participants may vote on FL model updates in a biased manner, either independently or collusively [147].

Moreover, the additional computational and storage demands of Blockchain lead to efficiency issues for BlockFL [159]. Energy consumption and new control costs arise from coordinating Blockchain operations with FL processes, as shown in Table 2. From the learning efficiency standpoint, empirical studies indicate that the convergence speed of BlockFL models slightly lags behind traditional FL [152], signaling opportunities for further optimization. Ongoing research is increasingly focused on enhancing the operational efficiency of BlockFL across various domains, which is anticipated to augment its practical utility significantly.

As PIoT applications become more widespread, the massive amounts of sensitive information used for training models pose significant challenges to privacy protection. In recent years, data security and privacy protection have garnered increased attention from researchers, particularly in relation to data generated during the sensing, communication, and computation processes of PIoT.

FL is at risk of data leakage when facing adversaries with an honest-but-curious server [8] or with Generative Adversarial Network (GAN) technology [37]. Although Blockchain can promote the development of decentralized and data-intensive applications [22], FL still relies on the honesty of miners as all raw data are public. Therefore, traditional FL and separate Blockchain technologies cannot satisfy the security and privacy requirements of PIoT scenarios. Hence, the BlockFL, which combines the advantages of FL and Blockchain, has become a new research direction to solve security and privacy issues in the PIoT.

A number of researchers have proposed different solutions for addressing the challenges of security and privacy in FL with the integration of Blockchain technology. Awan et al. [10] present a Blockchain-based PPFL model, which combines the FL framework with the decentralized trust of Blockchain to ensure privacy preservation. To achieve this, the authors enhance a variant of the Paillier cryptosystem to implement homomorphic encryption. Yin et al. [150] propose an FDC framework based on FL and Blockchain, which leverages multiparty secure computation technologies to ensure data security. Wang et al. [131] discuss the Security Parameter Aggregation Mechanisms in detail in their BlockFedML model. Furthermore, Ma et al. [72] propose a new group-based Shapley value computation framework that is compatible with secure aggregation in a Blockchain-based FL model. The approaches aim at addressing the privacy and security concerns in FL by integrating Blockchain technology and novel cryptographic methods.

The proliferation of IIoT has resulted in an exponential increase in the volume of data generated by devices equipped by various industries. The value of the data, because of the sensitive information it contains, has gained rise to concerns about data security. The leakage of IIoT data could result in significant financial losses for the company, as well as disruption and disorder within the industry.

Ensuring data security is a crucial factor in determining the utility of the IIoT model. Wang et al. [118] identify the security requirements for IIoT and investigate the advantages of integrating Blockchain technology into IIoT applications. In a separate study, Blockchain is leveraged in edge intelligence to optimize resource allocation in IIoT [163]. Additionally, FL has also been highlighted in IIoT applications [149].

To satisfy differential privacy, Geyer et al. [31] propose a method to conceal the contribution of each client during the training process. In the pursuit of safer data sharing, Lu et al. [70] build data models with BlockFL structures, where only FL-generated data models are shared by Blockchain. And thus, the model reduces the risk of raw data leakage and effectively protects data security. By using homomorphic encryption and secure multi-party computation, the authors ensure that the privacy of the raw data is maintained while enabling collaborative learning. Furthermore, Yazdinejad et al. [149] develop a block hunter framework based on cluster detection to automatically search for attacks and threat risks in BlockFL networks.

In IoV, practical models require large amounts of data sharing, which can include sensitive private information such as frequently visited addresses, real-time road conditions, driving routes, and driving preferences. Protecting privacy while participating in model training and sharing information with others is important and necessary in IoV applications [162], where the BlockFL framework could play an influential role.

Liu et al. [64] improve an optimized mask noise model upload algorithm for secure secret sharing of model parameters. The authors also introduce a two-stage Intrusion Detection System (IDS) utilizing the combination of FL and Blockchain in vehicles and roadside units to ensure data security and privacy protection in IoV. Chen et al. [18] propose a novel Byzantine-fault-tolerant Blockchain-based FL method named BDFL, which implements a publicly verifiable secret sharing scheme to address privacy concerns in IoV. The experimental results on actual datasets demonstrate the practicality of multi-object recognition while preserving privacy.

In the IoHT, a large amount of sensitive information poses a significant risk of privacy breaches. To address the issue and enable secure data sharing, BlockFL has been introduced as a promising approach for IoHT applications. Passerat-Palmbach et al. [91] propose a basic structure of Blockchain-orchestrated FL and identify six critical elements of privacy and security requirements in IoHT models:

The requirements provide a comprehensive framework for developing secure and privacy-preserving IoHT applications using Blockchain-orchestrated FL.

Based on the requirements, Połap et al. [95] develop a multi-agent system that divides specific medical tasks into agent units for parallel training of classifiers with FL and uses Blockchain to share and protect private data. Similarly, Aich et al. [1] design a BlockFL-based solution for the secure sharing of healthcare data to address the fragmented nature of personal medical data. However, the approach has only been theoretically analyzed without using practical applications yet.

El Rifai et al. [26] conduct experiments on a diabetes dataset to evaluate the effectiveness of their model, which utilizes BlockFL for secure knowledge sharing between medical centers while preventing attackers from accessing the raw records of the patients. The experimental results demonstrate the ability of the proposed approach to ensure data security and privacy protection. A hybrid Blockchain-based FL framework [101] has been tested in the context of COVID-19 clinical trials, which ensures the complete privacy of training data and supports reputation management, making it more relevant to current healthcare applications. Another privacy-preservation framework proposed by Singh et al. [117] also illustrates that the BlockFL technology can mitigate the risk of exposing patient medical data, creating a transparent and secure environment for data sharing and model training

Table 5 highlights models for IoT security and privacy in PIoT, IIoT, IoV, and IoHT. In PIoT, BC-based PPFL [10] focuses on privacy with enhancements like a Paillier cryptosystem variant, facing parallelism and non-IID data challenges. FDC [150] offers secure multiparty data computation with flexible access control, though with efficiency concerns. BlockFedML [131] provides secure aggregation with audit trails and encrypted communications but lacks experimental validation. For IIoT, BlockFL Data Models [70] enhance secure data sharing with differential privacy but struggle with utility optimization. Block Hunter [149] employs cluster-based anomaly detection for threat hunting, constrained by resource limitations. In IoV, a two-stage IDS [64] ensures safety in autonomous driving with a mask noise model, limited by resources. BDFL [18] aims at privacy preservation in autonomous vehicles using the HyRand protocol, facing scalability challenges. In IoHT, a multi-agent system [95] secures medical data processing with real-time capabilities, hindered by latency. A hybrid BlockFL [101] preserves privacy with encryption methods, yet with resource constraints. Secure Architecture in Smart Healthcare [117] maintains privacy in a distributed environment, dealing with latency and resource issues.

As shown in Figure 7, BlockFL models are making substantial strides in enhancing security and privacy across various IoT domains, addressing specific challenges and pertinent objectives. In PIoT, the emphasis is on preserving the privacy of vast amounts of sensitive consumer information, while in IIoT, the focus shifts to securing corporate data to ensure industry stability [63]. For IoV, the critical concern is secure information sharing to maintain the integrity of autonomous driving safety, and in IoHT, the models underscore the need to protect sensitive healthcare information, which is vital for maintaining patient trust and adhering to regulatory standards. Particularly in IoHT, where the stakes are incredibly high due to the involvement of personal medical data, BlockFL models are not only safeguarding privacy but also ensuring the security of data, which directly impacts the quality of healthcare services [119]. By leveraging the inherent properties of blockchain, such as immutability and decentralized consensus, along with the data localization of FL, these models foster innovation in healthcare technologies by enabling secure data sharing without compromising on privacy.

The PIoT model has to ensure the trust and availability of information, and services to enable its practical application. To avoid errors and losses in data transmission and communication, which can have catastrophic consequences (e.g., the risk of poisoning attacks), [12] has demonstrated the possibility of malicious workers manipulating model results by injecting poisonous data or tampering with training data.

Existing studies have been developed to tackle the challenge of information errors and losses during traditional FL training. Zhao et al. [166] design a method to address the issue by minimizing the influence of low-quality participants. VerifyNet [143] introduces a verifiable framework that verifies the integrity of the aggregated results of FL. However, the framework faces the problem of susceptibility to single-point attacks due to its reliance on the central server.

The combination of Blockchain and FL technologies improves the reliability of PIoT models because of the auditability and decentralization provided by Blockchain. Peng et al. [92] improve the VFChain system, enabling verification for the FL training process and recording verifiable proofs in the Blockchain. Preuveneers et al. [96] introduce an anomaly detection model into the FL process and utilize Blockchain technology to record its incremental updates. Kang et al. [47] propose a metric called “reputation” to support reliable-worker selection, ensuring data integrity and preventing tampering. By reducing the impact of adversarial data corruption, integrating Blockchain and FL technologies can improve the robustness and stability of PIoT models. The work in [127] applies the concept of reputation in a Blockchain-based fine-grained FL model to facilitate trustworthy collaborative training. The experiment in [120] shows that implementing the Blockchain in FL improves the performance when adopting various types of corruption to the dataset of the end-point adversary, including salt and pepper noise and circle occlusion. The FLchain scheme developed by Majeed et al. [73] outperforms traditional FL models in robustness as the provenance of data is auditable.

Real-world industries require reliable and stable IIoT models that can withstand environmental disturbances and attacks. However, data flaws are common in the FL process as local datasets are easily disturbed by environmental factors [36]. The usage of Blockchain for the underlying mechanism of the IIoT model can ensure the regular operation of the entire system so that the machines can work honestly and normally [112].

To detect device failures and attacks in IIoT applications, Zhang et al. [164] introduce an optimization of an averaging algorithm called CDW-FedAvg that calculates the distance between positive and negative class data. By combining the advantages of both FL and Blockchain, the developed approach reduces the impact of device failures and improves the stability of the IIoT system. Similarly, Qu et al. [99] develop a D2C paradigm for the IIoT model and a modified Markovian decision process to enhance performance when facing poisoning attacks.

Stability is a crucial advantage of industrial automation, enabling control and prediction of the operational status of machines. In Industry 4.0, researchers aim at enhancing the stability of IIoT models by combining FL and Blockchain technologies. Hua et al. [39] conduct experiments on heavy haul rail applications, replacing manual operation with intelligent control using a Blockchain-based asynchronous FL system. The simulation results demonstrated that the proposed BlockFL system effectively achieves stable and smart control in real heavy-haul rail applications.

The high mobility of vehicles in IoV introduces dynamically and rapidly changing environments, leading to crucial timeliness of decisions, as autonomous driving and intelligent transportation systems require vehicles to respond quickly to real-world situations. Therefore, improving the speed of vehicles’ model learning and information communication is a key issue in practical applications and a prominent topic in research.

To accelerate model learning and information communication in the IoV, Pokhrel et al. [94] develop a mathematical analysis to identify the delay in the BlockFL model, where participating vehicles share their on-vehicle machine learning model updates via Blockchain and cooperate to complete the FL process. The vehicles calculate the total end-to-end latency, including communication and consensus delays, and an online algorithm is proposed to adjust parameters in real time to minimize the model delay.

To improve the intelligence of vehicles, different models are needed to process, analyze, and respond to different application scenarios. Each model in the IoV requires diverse and vast data that is collected from the vehicles, the neighbors, and the roadside units. Due to the variability of the IoV, neighboring vehicles are in a constant state of flux, so vehicles in the IoV are often unfamiliar with their surroundings. It is, therefore, essential to evaluate the credibility of the data and identify any malicious attempts to compromise it.

Blockchain has been seen as an effective tool to integrate with FL as BlockFL to manage participants and improve system reliability [148] to address the trustworthiness issue in the IoV. PermiDAG model developed by Lu et al. [71] uses a hybrid Blockchain with a directed acyclic graph to perform asynchronous FL. The quality of shared parameters is verified to detect false information and malicious data. Kang et al. [45] introduce a reputation system to judge participant trustworthiness and design a distributed reputation calculation scheme for selecting trustworthy participants. The authors present a stable many-to-one matching model for task assignment to achieve a trusted win-win situation.

In the realm of IoHT, trust and reliability are paramount, with the sector demanding robust frameworks to guarantee data privacy and secure sharing of sensitive health information. This need has been accentuated by the challenges posed by the COVID-19 pandemic [107], where concerns about misinformation and malicious behavior have become rife. Moreover, the growing demand for personalized medicine presents new challenges to traditional healthcare systems, requiring innovations to accommodate individual patient needs [52].

In this context, the BlockFL model has gained attention for its potential to foster trust and reliability in healthcare data management. Samuel et al. have illustrated this by proposing a BlockFL-based infrastructure that not only enhances the dissemination of authentic COVID-19 information but also offers a robust infrastructure against security attacks [107]. Similarly, Moulahi et al. create a trusted Blockchain-based FL system capable of predicting diabetes risk, achieving significant accuracy in their results and demonstrating the resilience of the system against cyberattacks [82]. Some work focuses on the trustworthiness, accountability, and fairness of FL systems in IoHT. A BlockFL architecture is proposed in [68] design to improve these aspects by employing smart contracts and a fair data sampler algorithm. It is demonstrated to be feasible, enabling accountability and improving fairness in a COVID-19 X-ray detection use case.

Personalized precision medicine as a transformative approach to healthcare, which focuses on customizing treatments and therapies for individual patients based on their unique medical data [52], requires a high level of trust and reliability in managing sensitive health information, facilitated by advanced technologies like Blockchain and FL. Ali et al. have explored the transformative potential of Blockchain-enabled FL for precision medicine with a representative dataset of electronic medical records (EMRs) [4]. They emphasize the importance of trust and decentralized data sharing and demonstrate the feasibility of BlockFL by simulating an EMRs system. To address the need for personalized healthcare models and the challenges of potential risks, Lian et al. propose a Blockchain-based personalized FL system for IoHT with personalization layers to capture personalized features, which shows better results on heterogeneous medical data than a one-size-fits-all global model [62].

Table 6 shows the comparative analysis of BlockFL across various IoT domains for Trust and Reliability, which reveals distinct approaches tailored to the unique challenges of each sector. In the realm of PIoT, models like VFChain [92] prioritize verifiability and auditability to enhance data integrity, whereas IIoT solutions like the D2C paradigm [99] focus on resistance to sophisticated cyber threats, reflecting the critical need for robust defense mechanisms in industrial settings. For IoV, the BFL model underscores the importance of minimizing delays and ensuring data reliability to support the dynamic nature of autonomous vehicles [94]. In the healthcare sector, IoHT, the sensitivity of data in IoHT necessitates models that not only bolster system security but also address the pressing need for accountable and fair data handling, as demonstrated by the COVID-19 X-ray Detection [68] and Personalized Healthcare models [62]. These innovations collectively underscore the pivotal role of BlockFL in fortifying trust and reliability, with each domain benefiting from bespoke features that address their specific security challenges and operational demands.

BlockFL models are crucial in reinforcing trust and reliability within the IoT sphere, as shown in Figure 8. The models cater to the unique demands of each domain: PIoT focuses on data transmission accuracy, IIoT on correcting data flaws due to environmental impacts, IoV on timely decision-making for mobile units, and IoHT on safeguarding sensitive health data. Particularly in IIoT, where uninterrupted operation in harsh conditions is common, the reliability requirements highlight the importance of BlockFL models. The robust BlockFL frameworks counteract exploitation threats and ensure continuous industrial activities, playing a key role in protecting against economic and safety risks.

The practicality of the PIoT model is to enable intelligent PIoT devices and offer advanced services. To accomplish this ambition, the PIoT model should exhibit high accuracy and efficiency when executing tasks, which means optimizing algorithms and designing models for improving performance under the condition of limited resources.

Accuracy is a key metric for model evaluation in PIoT models. Existing works have demonstrated that BlockFL models integrating Blockchain and FL technologies outperform traditional FL and separate Blockchain in various tasks. Liu et al. [66] proposed the FedAC model, which combines asynchronous FL and Blockchain technologies, and achieved impressive accuracy rates of 98.96% in horizontal data distribution and 95.84% in vertical data distribution, outperforming the accuracy of its counterparts.

Efficiency is also an important metric for evaluating BlockFL models. To improve efficiency, Ramanan et al. [102] present the BAFFLE model by using smart contracts in Blockchain to coordinate the round delineation, model aggregation, and update tasks in FL. The model significantly reduces the computational cost of the model because smart contracts are computerized transaction protocols [122] that automatically execute the contractual terms. Feng et al. [27] develop two complementary policies to ensure efficiency, i.e., controlling the block generation rate and dynamically adjusting the number of training times in asynchronous FL. In ChainsFL [158], synchronous and asynchronous training are combined to improve the efficiency of the model.

In research and analysis, the existing studies assume that devices participating in model training have unlimited energy and ample computing power. However, in the real-world, industrial machines used in manufacturing often fall short of theoretical ideals. Due to the cost considerations, such equipment is subject to constraints on capacity, energy, communication ability, and other aspects. For instance, machines with limited computing power require more time to train and update the model, while those with poor wireless channel conditions take longer to transmit information. Therefore, it is essential to flexibly adjust model parameters based on actual conditions and enhance model performance under resource limitations.

To address the problem of resource constraints, Nishio et al. [89] develop the FedCS model, which selects suitable training participants. By excluding unqualified machines, as many participants as possible can join the training process under limited conditions, making it suitable for actual industrial applications. In addition to participant selection, adjusting other model parameters is also effective. Qu et al. [98] consider a range of factors, including communication, delays, and computation cost, to determine the optimal block generation rate in FL-Block, an autonomous FL system based on Blockchain.

Reducing energy consumption can be utilized to address the resource issues in BlockFL training. Lu et al. [69] improve a compression technique to reduce communication costs without sacrificing performance. The authors consider the instability and complexity of the network connections in the IIoT model, allowing machines to join or leave the training process more freely. Kang et al. [44] employ a gradient compression scheme to replace complete gradients with sparse but important gradients, effectively reducing communication overhead.

The dynamic nature of vehicular networks introduces a challenge in resource allocation. Despite the advancements in in-vehicle computing and communication technologies, there still exists a gap in achieving optimal solutions for model learning in theory. This is especially true in the case of BlockFL, where vehicles need to conduct multiple rounds of communication and require high computational power. Hence, exploring ways to adjust parameters effectively to meet the requirements under limited resources is an important research direction in IoV.

Chai et al. [16] improve a hierarchical FL algorithm that leverages Blockchain technology to include multiple ground chains and one top chain, resulting in reduced computation and sharing consumption. The experimental results show the effectiveness of the hierarchical structure. Pokhrel et al. [93] introduce a Blockchain-empowered FL system for drones in 6G networks aimed at disaster response systems. The authors focus on the impact of transmission parameters such as power and the number of miners on energy consumption through modeling and simulation that offer valuable insights and potential research directions for future work in this field. The negative impact of the energy limitation problem of drones on the service time is also discussed in a data collection BlockFL scheme [41].

Efficiency and resource management emerge as non-negligible concerns in IoHT, underscoring the need for solutions that optimize data processing and minimize energy consumption. The dynamic and distributed nature of IoHT devices, which collect vast amounts of patient data, poses significant challenges in terms of bandwidth usage, storage capacity, and computational load [161].

Traditional centralized data processing models often struggle with these challenges, leading to inefficiencies and potential risks. The application of Blockchain-empowered FL addresses these issues by decentralizing data analysis, thus reducing the need for data to be transmitted to a central server for processing. Lakhan et al. [56] introduce an FL-based Blockchain system to minimize energy consumption and delay in healthcare applications, showcasing the potential of BlockFL in meeting the stringent requirements of healthcare workloads with resource constraints. While Muazu et al. [83] demonstrate how edge computing, combined with BlockFL, can optimize resource management in IoHT, reducing computing costs while enhancing security and privacy. The above studies collectively underscore the transformative impact of BlockFL on IoHT, highlighting the role of BlockFL in achieving efficiency within the healthcare domain, especially with the rapid development of healthcare-based cyber-physical systems [67].

Table 7 presents models designed to boost efficiency in four domains. In PIoT, the BAFFLE model [102] enhances computational performance by segmenting parameter space, reducing costs but grappling with complexities in large models. BAFL [27] intertwines power with model utility for learning efficiency, though optimization remains a challenge. For IIoT, ChainsFL [158] uses a DAG-based network for better efficiency and scalability, despite resource limitations. FL-Block [98] leverages fog computing for cost-effective communication and consensus, yet faces stability issues. PAFLM [69] optimizes utility via edge network ML, with synchronization hurdles. In IoV, Hierarchical BlockFL [16] and BlockFL for Drones [93] focus on knowledge sharing and efficient consensus in IoT drones, contending with scalability in dynamic environments. For IoHT, FL-BETS [56] addresses task scheduling with risk quantification, balancing resource allocation and fraud vulnerability. Edge-empowered BlockFL [83] aims at resourcing optimization through mixed-model programming, improving allocation and reducing consumption, yet faces scalability issues.

BlockFL is instrumental in enhancing efficiency across IoT applications, with a notable impact in the IoV domain. As shown in Figure 9, BlockFL models in PIoT aim for high efficiency within limited resources to unlock advanced consumer services [3], while stakeholders and researchers in the IIoT focus on tackling cost-related performance gaps in manufacturing [28]. The IoV sector benefits from the ability of BlockFL to manage dynamic network resources effectively, optimizing real-time data processing for safer and more efficient transportation. Unlike its counterparts, BlockFL for IoV distinctly targets the swift allocation of resources amidst the dynamism of the network. This precision in IoV allows quick data handling and optimizes the decision-making process for safer and more efficient traffic systems [141]. In the realm of the IoHT, BlockFL is critical for managing the efficiency of distributed devices, ensuring swift data handling while conserving resources. The unique challenges of each domain are addressed through the collaborative model training in BlockFL, demonstrating the versatility and necessity of BlockFL in achieving efficiency and resource management capability in the IoT ecosystem.

Accurate judgments and correct decisions rely on the large amount and diversity of data. So increasing the enthusiasm of devices for participation in the model training is an effective action for optimizing the accuracy of PIoT models, which requires the model to have a reasonable incentive mechanism.

The Blockchain has shown its ability to provide an effective incentive mechanism based on the performance of participants. More and more recent research works have been done to implement Blockchain into PIoT applications, especially with the FL that can safely combine massive devices to train a model. Without requiring honest participants, Short et al. [116] offer rewards on a Blockchain network according to the quality of contributions in FL. Martinez et al. [79] propose an in-depth workflow to record and reward the contributions of participants. In the work of Kim et al. [51], Blockchain is used to separate participating users as nodes and induce them to join the FL efficiently.

Besides, in order to attract more participants to join BlockFL to improve data diversity and model performance, more work is devoted to designing more reasonable and attractive incentive mechanisms. An effective incentive mechanism combining reputation management with smart contracts is proposed by Kang et al. [46] to motivate high-quality devices to join the model learning process. Zhao et al. [167] design an incentive mechanism to award participants with a novel normalization technique. Weng et al. [139] propose a DeepChain framework with a value-driven incentive mechanism to force the participants to train the model following the rules. Kumar et al. [55] also develop a value-driven incentive mechanism to encourage the positive actions of the contributors by introducing Blockchain technology via Ethereum.

Introducing repeated competition for FL is also feasible [125] as rational participants want to maximize their profits. Also, based on the hypothesis of rational man, Xuan et al. [146] propose a double-layer FL platform based on Blockchain with an incentive mechanism to ensure that rational workers can gain the maximum benefit by remaining honest. Desai et al. [24] create a general Blockchain-based FL framework to detect and punish attackers automatically. An honest trainer [11] is presented to gain fairly partitioned profit, rewarding contributions, and punishing the malicious.

In the realm of IIoT, the diversity of training data plays a pivotal role in enhancing model performance, particularly in applications that demand high precision and reliability across various industrial sectors. The application of BlockFL emerges as a powerful solution, as it provides incentives for IIoT data owners, encouraging participation and thereby enriching the diversity and quantity of training data.

Recent research highlights the significance of incentive mechanisms in enhancing the diversity of training data for IIoT systems, addressing a critical challenge for the deployment of BlockFL. The blockchain and FL-based secure data-sharing scheme introduced in [144], incorporating model parameter validation and incentives into the consensus algorithm, directly addresses the challenge of data diversity by encouraging a broader participation base. Wang et al. have introduced an incentive mechanism for resource allocation in Blockchain-based FL, facilitating optimal participation by rewarding training and mining efforts [137]. The mechanism addresses the dual challenges of encouraging participation and managing resource constraints, thereby supporting the collection of diverse training data across IIoT devices.

The studies present innovative solutions that not only incentivize participation but also ensure the security and fairness of data contributions, crucial for the IIoT environment where data originates from a vast and varied array of sources. A Blockchain-based FL system, FGFL [30], focuses on a fair incentive mechanism, designed to attract high-quality workers and deter malicious actors by rewarding substantial and genuine contributions. Similarly, the Blockchain-enabled FL framework proposed by Witt et al. [140] aims at balancing contributions fairly, tackling the diversity challenge by ensuring that a wide range of participants are motivated to contribute their unique data. This framework uses Federated Knowledge Distillation with compressed soft-labels to promote honest participation through an incentive-compatible ecosystem.

In addressing the challenge of enhancing the diversity of training data within IoV, integrating BlockFL has proven to be a promising solution. The methodologies and frameworks developed in recent research focus on leveraging BlockFL to overcome data silos while promoting the diversity and quality of the data involved in training ML models.

For instance, frameworks like IoV-SFL [126] introduce secure and efficient data-sharing mechanisms, utilizing advanced encryption techniques and ML models to process diverse and heterogeneous data types while enhancing model performance. Fu et al. have incorporated a supervision game into a hierarchical Blockchain-supported FL architecture for autonomous driving, which demonstrates a significant stride in managing heterogeneous data within the IoV ecosystem [29].

Other studies emphasize the importance of incentive mechanisms to attract quality data contributions. Wang et al. have proposed BPFL, a Blockchain-based privacy-preserving FL scheme for IoV, enhancing Multi-Krum technology with homomorphic encryption [130]. Additionally, a reputation-based incentive mechanism is developed to encourage honest participation and data sharing. Similarly, BESIFL is introduced by Xu et al. [145], leveraging Blockchain for decentralization, integrating mechanisms for malicious node detection and incentive management, thereby improving FL performance by ensuring the participation of credible nodes. The paradigm underscores the importance of a secure and incentivized environment for handling diverse data and enhancing the overall efficacy of federated learning in IoV.

When discussing how to enrich the diversity of training data, the heterogeneity of healthcare data is a significant issue in the context of IoHT due to the diversity of medical equipment and the complexity of application scenarios. Because of the variability in data format and characteristics across different medical institutions, it is impractical to enforce a standardized structure for all data in IoHT. For example, Computed Tomography (CT) images may vary in size, pixel density, and data format, so addressing the heterogeneity of healthcare data is an important consideration for developing effective IoHT solutions.

To address the challenge of heterogeneity in healthcare data, Kumar et al. [54] improve a Blockchain-empowered FL model with the data normalization technique. The authors utilize capsule-network-based segmentation and classification to detect patterns of COVID-19 from various types of lung CT scans. By leveraging the Blockchain and FL technologies, the presented BlockFL model caters to the particularities of the IoHT.

Table 8 provides an assessment of BlockFL applications across various IoT sectors, highlighting their role in managing data diversity. DeepChain [139] in PIoT incentivizes nodes for collaborative training, ensuring fairness but facing synchronization issues. DAM-SE [146] promotes honesty with a cost-effective aggregation model, balancing strong data privacy against scalability and delay challenges. Blockfla [24] combats malicious intent by building trust indices, trading off with potential system overhead. FleChain [11] ensures equitable participation yet deals with model convergence. In IIoT, Secured Data-sharing [144] and FGFL [30] focus on fair compensation and attack resilience, respectively, each facing overhead and resource scalability issues. For IoV, BPFL [130] and BESIFL [145] use reputation-based and contribution-based incentives to enhance participation and data security, confronting node failure and efficiency impacts. In IoHT, COVID-19 Detection [54] employs a Capsule Network for secure data sharing in medical contexts, with limitations in resource availability.

Figure 10 delineates a structured analysis of the impact of data diversity on the effectiveness of BlockFL across varied IoT domains. Within PIoT, the framework emphasizes the critical role of incentive mechanisms in model training processes. In the realm of IIoT, the precision and performance of models are of utmost importance, thereby requiring the implementation of secured data-sharing schemes, maintaining data integrity, and improving operational efficiency. The IoV sector seeks to establish secure and incentivizing environments, aiming at attracting a rich variety of datasets and address the challenges of data silos. In the domain of IoHT, the focus shifts to tackling the heterogeneity of medical data, with sophisticated applications such as processing CT Imaging, developed to navigate the diversity of medical devices and the intricacies of healthcare application scenarios.

In the 5G/6G era, BlockFL is set to thrive due to unprecedented data speeds, lower latency, and higher reliability [106]. These advanced networks enhance real-time interactions between Blockchain nodes and FL participants, contributing to more efficient data exchange, model updating, and consensus. The increased data transmission speeds also aid in the quick synchronization of Blockchain ledgers and fast sharing of FL model updates, bolstering the scalability and efficiency of BlockFL systems. With the evolution of 5G and the advent of 6G technologies, however, BlockFL faces new challenges in maintaining efficiency and security. The ultra-low latency of these networks demands real-time data processing and decision-making in BlockFL systems [20]. For instance, autonomous driving requires millisecond-precise decision-making utilizing the low latency of these networks. Additionally, the higher data throughput of 5G/6G networks increases resource demands, necessitating more efficient BlockFL algorithms [152], especially in contexts like smart cities where vast sensor data needs to be managed without overwhelming network nodes.

As BlockFL navigates these technological advancements, it also encounters universal challenges in IoT scenarios, underscoring the need for tailored solutions in various applications. While each considered IoT scenario has its unique characteristics, they share foundational challenges in data security, privacy, resource limitation, scalability, and data diversity. All scenarios involve data collection, transmission, and processing, often with sensitive information, necessitating robust security and evolving privacy protection [9]. The limited computing power and energy resources of IoT devices pose scalability challenges as the IoT ecosystem expands. Moreover, data diversity in terms of device capabilities and communication protocols requires intelligent processing techniques for efficient BlockFL deployment [40]. These challenges vary significantly across scenarios. For example, IoHT involves health-related data requiring strict compliance with healthcare regulations [49], while IIoT might focus more on protecting industrial processes and proprietary information [78]. The degree of resource limitation and scalability requirements can differ. PIoT devices might be more constrained in terms of battery life and processing power [114], whereas IIoT settings might prioritize the scalability of solutions across vast industrial networks with varying computational capacities [142]. The type and level of heterogeneity can vary widely; IoV deals with mobility-related data and connectivity challenges unique to vehicular networks [84], while IIoT must accommodate a wide range of industrial equipment and operational technologies.

Table 9 summarizes potential technologies for future BlockFL development. In terms of security enhancement and privacy protection, the integration of encryption and secure computing technologies has been effective [10, 121, 150]. Combining BlockFL with various encryption algorithms and noise addition methods [154, 160] or multi-party security technologies can further enhance Blockchain-based FL models. Additionally, data processing techniques like compression [44, 69] and normalization [54], along with smart contracts [76] and sharding mechanisms [151, 165], are suggested to address resource limitations and data heterogeneity, enhancing the capabilities of BlockFL.

Furthermore, ongoing research is needed to delve deeper into combining Blockchain and FL. Current BlockFL models mainly employ Blockchain for aggregation in the FL process, with less emphasis on enhancing Blockchain through FL. Intermediate results in BlockFL, such as the quality of local models, could be utilized in the consensus calculations of Blockchain [70], thus reducing the costs of computational and communication resources. Exploring new consensus methods and smart contract technologies based on BlockFL presents promising avenues for further development. This research should also address issues typically associated with Blockchain, such as the collusion of the miners and the challenges of hybrid-Blockchain structures [134, 157].

The integration of BlockFL across different IoT applications presents unique challenges due to the distinct nature of each field. Table 10 summarizes these challenges and their solutions. In PIoT, the complexity arising from personalized smart services is addressed using transfer and split learning technologies to enhance personalization and privacy. IIoT leverages transfer learning to improve intelligent collaboration, boosting efficiency and cutting costs. For IoV, high latency is tackled through online and continuous learning technologies, enhancing safety and traffic flow. In IoHT, the focus is on handling sensitive medical data securely, utilizing consortium Blockchains and split learning for secure data handling and collaborative research. This table encapsulates the diverse challenges and innovative solutions tailored to the unique needs of different IoT scenarios.

In large-scale PIoT and IIoT, collaborative intelligence and meeting personalized needs have emerged as a new research direction. As existing studies have focused on optimizing a single task, the increasing demand for multitasking collaboration and cooperation requires a complex system to analyze and coordinate the relationship and connection of multitasking and multi-objective. To enable intelligent coordination, the models should explore transfer learning technology and other related novel technologies in combination with BlockFL models. In particular, the process of industrialization requires more consideration of the costs of large-scale implementation.

In IoV, ensuring the stability of the system in high-speed movement is a crucial research direction. A large number of vehicles in the IoV application scenarios are constantly moving at high speeds and changing positions in real time, which poses a considerable challenge to the stability and reliability of the network and connection. To address this issue, researchers can increase the calculation effectiveness, reduce model delay, and consider optimizing and innovating BlockFL models by imitating online learning algorithms. Moreover, future vehicles in 6G are expected to support cross-domain communication across the ground, underwater and air [33], so stability in combination with new devices and technologies, such as over-the-air computing, should also be taken into account [168].

In IoHT, permission and identity management of participants are critical challenges due to the high sensitivity of medical data. Consortium Blockchains are more suitable for implementation in IoHT, with the high professional knowledge required by participants to analyze and manage medical-related data. The involvement of medical organizations can make the management of IoHT models highly controllable and convenient, and multiple participation can improve the accuracy and other performance of IoHT models. Thus, researchers should explore how to incentivize participation in BlockFL while considering the problem of membership management.

In addressing future research on BlockFL, it is imperative to broaden the scope to include various frameworks of distributed learning, such as split learning, transfer learning, and continuous learning. This expansion is crucial for offering a holistic view of the distributed learning landscape, enabling the identification of synergies and potential integrations that could further enhance the capabilities and applications of BlockFL across diverse IoT scenarios.

Integrating learning frameworks like Split Learning, Transfer Learning, and Continuous Learning with BlockFL offers the potential to leverage the strengths of both FL and Blockchain technologies while addressing their respective weaknesses. By doing so, we can create a learning ecosystem that is robust against data privacy issues, adaptable to new data, and capable of continuous improvement without centralized data storage.

Split Learning is a distributed learning framework that divides a neural network model between client and server sides [105]. Clients compute with local data and send intermediate results to a server for further processing. This method, combined with BlockFL, enhances privacy and efficiency. Clients handle initial training stages and only transfer intermediate results to a Blockchain-based server, which aggregates them securely and updates the Blockchain ledger with the enhanced model. This integration with BlockFL offers scalable, privacy-preserving solutions in IoT environments, benefiting from the security and transparency of Blockchain.

Transfer Learning applies knowledge from one domain to solve problems in another and is particularly useful in PIoT and IIoT within BlockFL contexts [138]. By incorporating it into BlockFL, pre-trained models on Blockchain nodes can be refined by new participants using their data, thus reducing training time while maintaining privacy. This approach also facilitates cross-domain applications, allowing knowledge transfer from one IoT sector to another (e.g., from IoHT to IoV), accelerating intelligent system deployment across varied IoT ecosystems.

Continuous Learning focuses on systems that learn and evolve over time, accumulating and adapting to new data while retaining previous knowledge [90]. Incorporating this into the BlockFL framework could enhance adaptability. The approach involves regularly updating Blockchain models with insights from client data. Clients contribute to ongoing learning, facilitating continuous model evolution, which ensures data integrity and lineage for auditing purposes. In dynamic IoT settings, this enables BlockFL systems to adapt to new patterns and changes, maintaining relevance and effectiveness in applications such as IoV and IoHT, where continuous learning and updating are essential.