As IoT ecosystems create immense data streams, FL offers an effective way to manage and analyze them. To address the constraints of centralized data handling in IoT networks, the community has released multiple open-source FL platforms [9]. By analyzing input streams directly on devices and federating results, these frameworks support privacy preservation and low latency [10]. Yet, IoT-focused FL implementations still struggle with issues of robustness, efficiency, and security. To mitigate these problems, novel solutions have been introduced, such as blockchain-enabled asynchronous FL scheme that verifies information accuracy, avoids dependence on vulnerable nodes, and boosts system productivity [11]. An additional approach combines blockchain with federated learning to obscure shared training parameters and enhance privacy protection in IoT environments [12]. To refine FL operations in IoT scenarios, several frameworks have been proposed. One such method, FedEAFO, optimizes performance by simultaneously reducing parameters and incorporating local updates, thereby balancing the trade-offs across data processing, connectivity, and accuracy [13]. For image recognition tasks in IoT, researchers have proposed a federated unsupervised learning approach capable of managing the difficulties created by non-IID logs among diverse clients [14]. SemiFL further integrates centralized and federated paradigms to support large-scale IoT systems, with notable benefits in settings that include computation restricted sensor devices [15].

AsyFed addresses device heterogeneity and resulting stragglers by clustering clients into gears based on comparable training performance and introducing a T step protocol that reduces the contribution of lagging gears [16]. In IoT environments, a streamlined hybrid federated learning framework employs blockchain-based smart contracts to ensure authentication, distribute models, and maintain trust [17]. FL has shown potential as a resource-efficient replacement for centralized intrusion detection in IoT environments. Studies employing shallow artificial neural networks and different aggregation methods confirmed that FL based IDSs can achieve competitive prediction accuracy, reliable response behavior, robust sensitivity, and solid F1-score [18]. To better preserve confidentiality with digital medical systems, studies have merged privacy-enhancing technologies and blockchain with federated learning. These combined methods tackle critical technological requirements for healthcare data protection, such as blockchain-enabled model repositories, protected aggregation procedures, and immutable gradient uploads [19]. Current FL frameworks for IoT show strong potential in addressing distributed data handling, privacy protection, and rapid model training. Yet, ongoing research continues to refine these systems, aiming to boost efficiency, robustness, and security while aligning with the evolving needs of IoT applications across various industries [20].

As a cooperative learning paradigm, FL supports the protection of user data, yet it simultaneously opens the door to specific security threats that malicious entities can leverage [21]. A significant risk arises from model poisoning, in which hostile participants manipulate local contributions to corrupt the shared model and disrupt its consistency [22]. Among the most severe risks in FL is model poisoning, in which attackers alter personalized updates to damage the shared model’s effectiveness. An advanced threat called the Sybil collusion attack has been identified in IIoT-FL setting. In this method, adversaries inject malicious updates by combining label flipping techniques with the creation of multiple Sybil nodes. This increases the likelihood that corrupted updates will be accepted during aggregation, leading to misclassification of targeted samples while the accuracy on other classes remains largely unaffected [23]. This approach introduces a verifiable incentive mechanism designed to encourage collaboration among decentralized participants while addressing persistent challenges in federated learning.

To mitigate poisoning threats, the DeMAC framework has been proposed as a tailored defense strategy. It leverages a metric known as GradScore, which assesses the gradient norms of client updates, thereby distinguishing malicious actors from legitimate contributors during training. A key advantage of this approach is its ability to automatically detect adversarial activity without manual parameter tuning, and empirical studies show it significantly reduces attack success rates across diverse poisoning strategies [24]. Security architectures supported by blockchain [25]. Homomorphic encryption-based federated learning models provide another method for achieving privacy preservation [26]. Despite existing efforts, it is still difficult to reconcile privacy; studies may explore enhanced federated learning architectures that tackle both privacy and security challenges concurrently [27,28].

Federated learning reduces data exposure, but it still depends on a central server, which poses issues with poisoning, integrity, and trust. FL-only systems struggle to confirm whether client updates are accurate or unaffected, and trust-scoring methods often lack robustness. However, blockchain offers decentralized auditability and immutable validation, but it does not provide efficient model aggregation, customization, or large-scale distributed learning. We combine the benefits of FL and blockchain by doing the following: Blockchain ensures verifiable, tamper-resistant update provenance and transparent incentive mechanisms, while FL minimizes data transfer and permits scalable, privacy-preserving training across heterogeneous IoT devices. This combination is becoming more and more popular in research on trustworthy FL [29] and decentralized federated learning frameworks that stress auditability and transparency [30].

This survey presents a succinct taxonomy based on attack surface, attacker type, and target layer because security risks in blockchain-assisted federated learning are still dispersed across layers. Gradient-based inference, which can extract secret data from shared updates [31], backdoor insertion, where hidden triggers are implanted during training, and Byzantine poisoning, where corrupted updates deteriorate model integrity [32], are important attack surfaces. Even in lightweight systems, dangers like ledger-state manipulation and consensus manipulation exist at the blockchain layer [33]. These risks target IoT devices, the FL aggregation pipeline, or the blockchain ledger and come from external adversaries, Sybil-style attackers, or malicious insiders. This cohesive framework replaces previous disjointed explanations and explains why blockchain-only or FL-only defenses are insufficient. Blockchain is increasingly recognized as a vital tool for boosting cybersecurity in diverse industries such as retail. Its immutable and distributed architecture provides a strong basis for mitigating prevalent security risks. Through distributed ledger mechanisms, blockchain also facilitates decentralized data governance while reinforcing integrity and trust [34]. The application of blockchain within retail settings has introduced stronger mechanisms for supply chain transparency, secure transactions, and reliable data handling. Because of the technology’s traceability features, product sources can be verified, which lowers fraud associated with counterfeit goods and improves vendor management. Moreover, blockchain enabled transaction frameworks, together with fraud monitoring solutions, add an extra shield of trust to financial operations in this sector [35]. Blockchain excels in protecting data by applying privacy controls, enforcing consent protocols, and maintaining permanent, unalterable records. Such capabilities are crucial for industries that manage sensitive datasets, including medical services. To advance biomedical security in this area, the BDL-IBS (Blockchain and distributed Ledger based improved Biomedical security) solution integrates blockchain with distributed lrdger systems, delivering enhanced safeguards for medical data [36]. Although blockchain first emerged in the financial sector, its applications have since broadened into multiple domains, including supply chain management, IoT, governance, and manufacturing. In recent years, adoption has grown steadily across these areas, with the banking industry witnessing some of the fastest advancements [37]. Despite its advantages, blockchain continues to encounter obstacles including scalability limitations, security vulnerabilities, and unresolved regulatory issues. Addressing these challenges is essential to achieving large scale deployment of blockchain based systems [38]. Even with existing hurdles, blockchain stands out as a powerful tool for advancing cybersecurity practices in multiple industries. By enabling secure, transparent, and distributed frameworks, it strengthens resilience against cyberattacks. Ongoing research and adoption are likely to foster novel blockchain based protection strategies in the coming years [39]. Table 2 explains how the main attack types in federated learning relate to the defence measures in blockchain-enabled frameworks. The table illustrates how security against model poisoning, Sybil attacks, inference assaults, and Byzantine behaviour is strengthened by blockchain characteristics like immutable logging, decentralised identity management, and transparent validation. The particular security benefits offered by blockchain-based federated learning under various attack scenarios are made clear by this mapping.

As a collaborative learning method, FL allows devices or organizations to train a common model collectively without transmitting their sensitive source data [51]. This method allows training to take place across multiple distributed nodes while ensuring robust safeguards for both privacy and data security [52]. In federated learning, computation is divided between edge devices and a coordinating server. Each device processes its local data and produces updates, which are then merged at the central node through an aggregation step [53]. A key difficulty in federated learning arises from the presence of non-independent and identically distributed datasets among clients, leading to heterogeneous learning contexts [54]. In response, tailored FL methods have been developed to produce client specific models, ensuring that the learning outcomes better reflect diverse user demands [55]. The combination of federated learning and blockchain is also under study, aiming to provide more secure, decentralized, and incentive driven collaborative systems [56]. To put in briefly, federated learning offers a powerful framework for safe AI development, particularly beneficial in domains like healthcare and financial services, where data transfer is prohibited [57]. As federated learning progresses, efforts are being directed towards enhancing communication efficiency, optimizing model aggregation techniques, and reinforcing defenses to ensure stronger and more scalable systems [58]. The design of effective FL systems plays a vital role in supporting algorithm deployment and overcoming challenges tied to accuracy, resource utilization, and data confidentiality [59].

While FL offers a valuable pathway for secure and distributed predictive learning, it continues to encounter challenges in safeguarding privacy and maintaining data reliability. The decentralized setup leaves room for malicious actors to exploit framework weaknesses [60]. Backdoor attacks and malicious exploitation pose a significant threat to FL, as they allow attackers to alter the shared model by injecting harmful inputs or poisoned updates [61]. Privacy can also be compromised through inference attacks, which make it possible for attackers to deduce original data by analyzing the parameters communicated during training [62]. Because of its distributed design, FL is exposed to several other attack surfaces that malicious actors can exploit to target the global model. Paradoxically, while FL aims to enhance confidentiality, studies have demonstrated that it is not fully immune to privacy related attacks [63]. This paradox highlights the necessity of integrating stronger defense strategies into FL frameworks. Researchers have proposed countermeasures applied at different stages prior to aggregation, during the process, and after model updates are combined [64]. Researchers are examining approaches like encrypted computation, privacy preserving anonymization, and blockchain assisted models to reinforce the confidentiality and resilience of FL [65]. To effectively tackle the privacy and security challenges in federated learning, comprehensive and multi-layered strategies are required. Future frameworks should embed the core strength of FL: its ability to protect sensitive data [66]. Future research should focus on balancing privacy, model accuracy, and stronger security.

Table 3 compares differential privacy, homomorphic encryption, and hybrid privacy-preserving models for IoT-based federated learning. The comparison shows that while DP-based methods are lightweight, they suffer from accuracy loss, whereas HE-based solutions offer stronger confidentiality at the cost of higher computation and latency. Hybrid and adaptive approaches attempt to balance these trade-offs, motivating the need for intelligent coordination in the proposed Hyper-Learning Framework.

Since privacy-preserving strategies in FL are frequently given inconsistently, this study groups them around two fundamental methods: Homomorphic Encryption (HE) and Differential Privacy (DP). Although DP can decrease accuracy under non-IID data, it is lightweight enough for IoT devices and adds calibrated noise to updates. HE prevents sensitive gradients from being exposed and maintains model changes encrypted throughout aggregation, but it adds significant computation and communication overhead and is mostly appropriate for gateways or edge servers. According to recent IEEE research, combining HE at aggregation nodes with DP at the device level can balance resource restrictions and privacy strength in heterogeneous IoT environments. The best times to use each method in blockchain-enabled FL systems are made clear by this methodical comparison. Despite the fact that FL employs a number of privacy-preserving techniques, their discussion in the literature is frequently dispersed and lacks a comparison framework. In order to solve this, we take a narrow emphasis on two popular methods: Homomorphic Encryption (HE) and Differential Privacy (DP). By adding calibrated noise to gradients prior to upload, DP protects user data and provides low computational overhead appropriate for IoT devices with limited resources. However, under non-IID settings, it increases accuracy loss [67]. HE is more suitable for gateways and MEC nodes since it allows encrypted model updates to be aggregated without decryption, offering high confidentiality assurances but imposing a substantial computational and communication expense [68]. According to recent research, hybrid designs can balance accuracy and overhead in heterogeneous IoT installations by combining DP at the edge with HE at intermediate nodes to provide layered safety [69]. This comparative viewpoint makes the privacy design decisions more understandable and in line with the real-world limitations of blockchain-assisted FL.

In the proposed Hyper-Learning Framework, the selection between Differential Privacy (DP) and Homomorphic Encryption (HE) is guided by a DRL-based control mechanism. The DRL agent observes device-level states such as computational capacity, energy availability, latency constraints, and communication reliability. Based on these observations, resource-constrained IoT devices are dynamically assigned DP-based protection to ensure lightweight privacy preservation, while more capable nodes such as gateways and edge servers employ HE for stronger confidentiality during aggregation. This adaptive selection enables a balance between privacy strength and system efficiency under heterogeneous IoT conditions [80]. By warping shared updates, Differential Privacy (DP) mainly thwarts inference-based attacks like membership inference and gradient inversion. Nevertheless, malevolent clients are still able to deliver poisoned model modifications thanks to DP. Conversely, Homomorphic Encryption (HE) protects against honest-but-curious aggregators by allowing encrypted aggregation, but it does not reduce poisoning, Sybil, or backdoor assaults. Because of this, neither DP nor HE by itself provides total protection against all attack surfaces, highlighting the need for other procedures like blockchain-based verification and trust management [81].

The use of blockchain within federated learning strengthens guarantees of trust, safety, and fairness. Decentralized consensus and validation protocols address potential weaknesses during training, particularly when dealing with unreliable or adversarial nodes. Through tamper-proof logging of model updates and user activities, blockchain ensures both integrity and privacy in distributed learning environments. With immutable tracking of model changes and participant actions, blockchain becomes a key enabler for secure and confidential collaborative learning [82]. By requiring organised involvement and public update validation, blockchain might lessen the influence of non-IID data, which is still a major problem in FL. In order to prevent excessive bias from highly skewed nodes, smart contract-based selection can provide preference to clients with trustworthy or balanced data [83]. Immutable on-chain logging aids in identifying aberrant gradient variance and local drifts in diverse environments [84], while blockchain-based reputation scores deter updates that consistently deviate from expected gradient behaviour [85]. These processes stabilise aggregation and enhance convergence among many IoT devices, even when blockchain does not completely eradicate non-IID distributions. While blockchain-assisted mechanisms such as reputation-based client selection and immutable update logging can alleviate the impact of non-IID data, they do not eliminate it entirely. In highly skewed data distributions, trusted clients may still contribute biased updates, and reputation scores may converge slowly in dynamic IoT environments. Moreover, blockchain-based validation increases latency and may limit rapid adaptation when data distributions shift abruptly. These limitations indicate that blockchain can mitigate, but not fully resolve, non-IID challenges, especially under strict resource and real-time constraints [64].

A key strength of blockchain lies in its immutability, ensuring that model updates remain permanent once committed, thus safeguarding the system from manipulation attempts. This feature provides strong guarantees in multi-party settings where stakeholders may not fully trust one another and where transparent auditing is necessary. Through consensus driven verification, blockchain enhances the protection and trustworthiness of federated learning systems. Given that FL relies on distributed clients who share only model updates, such mechanisms are critical for securing the joint model-building process. The Table 4 summarizes the recent works in blockchain based federated learning.

Through an immutable, verifiable record of model changes and contributor data, blockchain prevents falsification and markedly improves security guarantees [86]. The permanence of blockchain entries ensures that stored model updates cannot be rolled back, protecting the system from manipulation during training. This property proves critical when contributors may not fully trust one another or when transparency in auditing is required [87]. Ensuring the uniformity of operations in federated learning is a concern, and blockchain offers a solution. By leveraging smart contracts, it becomes possible to automate the enforcement of rules, making sure every participant abides by the established protocols [88]. Such a mechanism secures the correctness of the trained model and acts as a barrier against malicious actions like tampering with updates. Moreover, the decentralized validation offered by blockchain allows independent nodes to confirm the authenticity of updates before they affect the overall system [89]. Apart from enhancing protection and verifying the integrity of shared models, blockchain provides a foundation for transparent contribution tracking and fair incentive allocation among participants.

Ten representative studies that show how blockchain-enhanced FL has developed in various IoT scenarios are summarised in Table 5. Our explanation of safe participant authentication in Section 3.3 is reinforced by BDIM’s identity-management focus whereas the claims of cryptographic verification, tamper resistance, and trust formation are directly supported by FabricFL and LPBFL. The hierarchical UAV-IoT frameworks and RL-FL-BC are in line with Sections 5 and 5.1, showing how reinforcement learning enhances resource allocation, adaptive coordination, latency handling, and blockchain and FL. The paper’s focus on privacy protection, robustness, and non-IID data management is reinforced by edge-centric and multi-task learning contributions, such as decentralised FL over mesh networks and differentially private HDT systems. The scalability, security, and reliability issues examined in Sections 4 and 6 are substantiated by DAG-based blockchain for IIoT and multisignature-assisted FL. Lastly, our system-level discussions on mobility-aware optimisation and energy-efficient training directly relate to UAV-assisted MEC offloading. When taken as a whole, the works in Table 2 demonstrate the technical reasons for moving closer to the suggested Hyper-Learning Framework and confirm the main points made throughout the survey.

The verifiable nature of blockchain records allows the establishment of fair payment schemes, reinforcing trust and accountability among FL participants [90]. It is important to note that integrating blockchain and FL is not straightforward, since transaction costs and bottlenecks may arise, requiring thorough study in environments with constrained devices [91]. While blockchain enhances transparency, mishandling its design may put privacy at risk. Storing sensitive data or participant identifiers directly on an immutable ledger could inadvertently expose confidential information [92]. By combining blockchain with federated learning, researchers aim to build systems that are more secure, transparent, and dependable. Blockchain’s ability to keep permanent records, apply smart contracts automatically, and manage rewards fairly makes it a strong complement to FL. In the future, this integration is expected to produce hybrid models that offer both resilience and efficiency in collaborative learning [93].

The combined table and graph demonstrate that FedAvg stays close with little loss while centralised training provides the best accuracy. While expected decreases occur, more decentralised techniques like FedDec and MeshFedAvg continue to work reliably across datasets. Accuracy trade-offs are introduced by privacy and security-enhanced FL techniques, however these are greatly mitigated by optimized variations. The best adaptive performance is attained by RL-based FL, which frequently matches or surpasses baseline FL. Overall, the combined findings show that federated learning can continue to be precise, safe, and scalable, and the additional quantitative comparison enhances the suggested framework’s technical validity.

In blockchain based IoT applications, consensus mechanisms play a crucial role, and different approaches are being assessed for their performance and applicability. Notably, three well known algorithms such as custom proof-of-work, PBFT, and two valued consensus have delivered promising results in IoT settings by achieving consensus in less than one second [102]. In IoT environments, legacy consensus approaches like Proof of Work are limited by the modest processing strength of IoT nodes. Since PoW is highly computationally intensive, it proves unsuitable for devices with restricted energy and processing resources [103]. Addressing this concern, researchers have introduced more appropriate consensus protocols for IoT systems. Proof of Stake, in particular, has gained attention as it minimizes computational demands while significantly improving energy efficiency relative to PoW [104]. Practical Byzantine Fault Tolerance (PBFT) stands out as a reliable consensus framework for IoT, demonstrating compatibility with devices that have limited computational capacity [105]. The convergence of blockchain and IoT creates a dual scenario in which blockchain strengthens trust and security within IoT networks, the limited resources of IoT nodes highlight the necessity for optimized, resource efficient consensus mechanisms [106]. As innovations continue, specialized consensus protocols for IoT will likely evolve, crafted to overcome device constraints yet maximize the benefits blockchain contributes to distributed systems [107]. The summary of consensus mechanisms in IoT has been presented in Table 6. The degree to which the consensus techniques in Table 6 are appropriate for IoT environments varies significantly. Although they provide robust security, PoW and PoS are too hefty for low-power systems. When the number of nodes increases, PBFT and related protocols perform badly, but they are effective in small, permissioned groups. Although they can be less resilient to attacks, IoT-oriented and DAG-based techniques lower delay and energy consumption. The requirement for an adaptable, learning-driven consensus layer in the Hyper-Learning Framework is highlighted by the fact that no current architecture successfully combines trust, responsiveness, energy efficiency, and widespread participation.

Integrating blockchain within IoT networks faces major obstacles because of the restricted capabilities of IoT devices. Constraints in computing resources, limited memory, and high energy demands often hinder the efficient execution of blockchain-based operations in such environments [108]. Since IoT hardware generally offers minimal computational power and limited storage, these devices struggle to participate in conventional blockchain networks that demand intensive resources for consensus and mining [109]. The challenge is most apparent in edge nodes with limited capacity, where the high algorithmic load of blockchain proves difficult to manage [110]. Several solutions have been put forward, including the development of streamlined blockchain frameworks and mechanisms for balancing workloads across devices with varying resource availability [111]. A creative solution introduced in literature is the sliding window blockchain, where earlier block data supports the hashing of subsequent blocks, coupled with a simplified proof of work mechanism, making it more feasible for IoT use cases [112]. A different solution leverages the Proof of Authority mechanism in Ethereum which has been applied to show the viability of blockchain within IoT networks by mitigating both delay and scalability issues [113]. Overall, the combination of blockchain and IoT opens doors to improved decentralization, transparency, and security, but faces obstacles due to the limited resources of IoT hardware. Researchers are focusing on developing efficient blockchain and consensus mechanisms tailored for such environments ensuring the feasibility and scalability of blockchain driven IoT ecosystems [114].

Energy efficient allocation methods combined with dynamic node selection are instrumental in refining wireless network operations, notably in varied and device to device communication scenarios. Within heterogeneous networks, applying these techniques leads to marked gains in energy efficiency [115]. Attention is given to power reduction mechanisms and conflict mitigation for co-channel Single-RAT HetNets, underlining the necessity of flexible, energy optimized resource control frameworks [116]. An energy conscious staged allocation model for HetNets is proposed, incorporating fractional frequency reuse to both conserve resources and ensure stable connectivity at the network boundary. Through adaptive center edge partitioning, time slot planning, and optimized power management, the system secures notable improvements in efficiency and lower outage probability.

Leveraging direct D2D connectivity creates opportunities to implement power optimized strategies in wireless networks [117]. Research highlights that embedding D2D communication into dynamic TDD frameworks can boost energy efficiency by coordinating mode choice, slot allocation, and power usage. Such designs demonstrate improved spectral performance and reduced energy consumption over standard cellular methods [118]. It presents a new framework that improves D2D energy performance in HetNets by dynamically determining communication modes via fuzzy clustering. It further highlights that intelligent, dynamic selection of nodes within Cloud RANs is key to achieving better energy efficiency [119]. It highlights an integrated framework for heterogeneous C-RANs that unifies RRH activation, user association, and resource scheduling. Using a greedy activation policy and channel-driven pairing, the scheme achieves significant energy efficiency improvements. Ensuring network performance further requires dynamic node selection and energy optimized designs. Extending these methods to HetNets D2D links, and C-RANs strengthens both spectrum efficiency and overall performance. Additional measures such as sleep/wake cycles, interference reduction, and dynamic resource distribution contribute to building more sustainable and high performing wireless networks [120].

In FL, DRL serves as an effective approach to improve resource allocation, particularly in wireless environments where energy and bandwidth are limited. DRL-based strategies have been designed to identify optimal UAV placement and manage resource distribution in UAV-assisted networks. By accounting for the limited energy harvested by user devices, these methods help ensure the long term sustainability of FL operations [125]. The goal of this model is to improve the long-term effectiveness of federated learning systems in environments with limited resources, where obstacles include constrained bandwidth, harvested energy, and UAV power budgets. Energy constraints are transformed into a deterministic structure through the use of stability based optimization, which makes it possible for a DRL-based Markov Decision Process to direct effective system operation. There is room for more balanced designs, though, as the majority of the current research still focuses on either single control or fully distributed DRL approaches [126], A hybrid system known as federated deep reinforcement learning has been suggested to provide performance close to centralized training while decreasing information exchange and user privacy protection in wireless networks. In this setup, a central unit coordinates training but limits the scope of data sharing, making it highly suitable for federated learning environments. DRL-driven adaptive resource allocation further improves this approach by optimizing energy, bandwidth, and device deployment under restricted conditions, all while preserving privacy. Looking ahead, advancing specialized DRL algorithms tailored for secure and efficient FL remains an important direction for future research [127].

Because edge devices have limited battery capacity, compute power, and communication bandwidth, energy economy is a significant difficulty in federated learning for IoT systems. Energy consumption is greatly increased by frequent model updates, blockchain-related validation processes, and repeated local training, especially in large-scale and heterogeneous IoT implementations. Recent research focusses on energy-aware federated learning techniques that both maximise accuracy and resource utilisation to address these issues. By adjusting client participation, transmission power, update frequency, and aggregation intervals based on current device and network conditions, deep reinforcement learning has become a useful tool for dynamically managing energy consumption. IoT devices with limited resources can selectively participate in training thanks to this kind of adaptive control without compromising the overall performance of the model. By synchronising compute offloading, communication time, and device activation, DRL-based scheduling algorithms further enhance energy efficiency in edge-assisted and UAV-enabled IoT systems. These methods retain a reasonable convergence speed and learning accuracy while cutting down on pointless transmissions and avoiding overloading low-energy nodes. By moving intensive processing away from devices with limited resources, hierarchical FL architectures enabled by MEC nodes also aid in balancing energy consumption. In general, adaptive, context-aware techniques that concurrently take into account device heterogeneity, network dynamics, and learning objectives are needed for IoT-specific energy optimisation in federated learning. A viable technique to accomplish sustainable FL operation in IoT systems without sacrificing accuracy or scalability is through DRL-driven energy management.

While combining blockchain with FL strengthens data security and trust in IoT networks, deploying such systems at scale remains difficult. Resource constrained devices struggle to participate fully, and differences in hardware, connectivity instability, and communication delays hinder synchronous training. These factors collectively slow convergence, reduce model accuracy, and impair overall efficiency across heterogeneous IoT environments.

Managing resources and communication remains a major challenge for blockchain enabled FL in IoT networks. Systems can experience sharp performance drops under fluctuating network conditions and heterogeneous device capabilities. Adaptive mechanisms that can dynamically balance training loads and optimize communication schedules are essential. Additionally, blockchain integration imposes extra energy and computational demands, which can overwhelm low power devices. The choice of block sizes leads to frequent validation cycles, increasing energy usage and communication overhead.

In blockchain supported federated learning, energy consumption is a critical concern. Even supposedly low power consensus protocols like PoS still require significant computation, and multiple rounds of model training amplify the energy burden. Most current models neglect energy efficiency in key operations such as block generation, participant selection, and aggregation, resulting in fast battery drain and reduced sustainability for IoT devices. On the security side, FL remains exposed to threats including model poisoning, Sybil attacks, and privacy leakage. Malicious participants can manipulate model updates or incentive systems, and existing verification or reputation mechanisms often fall short in scalability and resilience against collusion. Furthermore, static approaches to selecting nodes for block creation and model aggregation increase vulnerability to rogue participants and compromise the integrity of the global model.

The integration of blockchain in federated learning creates extra strain on energy use and computational efficiency, which limits its suitability for constrained IoT systems. Among the most decisive issues is block size control. Poorly managed block sizes affect the overall network by influencing latency, throughput, and storage capacity. Small block sizes shorten consensus intervals but lead to higher communication load and energy costs, while large blocks delay consensus due to increased propagation delays. Beyond block size, the energy demand of consensus itself is a key obstacle. Even energy optimized mechanisms such as Proof of Stake still consume considerable resources. Since federated learning relies on frequent training and communication, the cumulative energy efficiency considerations in block production, aggregation design, and the process of selecting participating nodes.

Network stability is difficult to maintain since sensor-driven and mobile IoT devices often experience quick energy depletion. Security risks add another layer of concern. While blockchain offers some protection against tampering, federated learning is still exposed to threats such as data inference, Sybil behavior, and poisoned model updates. Attackers may deliberately inject false contributions or manipulate the incentive structure. Approaches like verification checks and reputation models have been proposed, but they typically fail to scale and are ineffective when multiple attackers act together. Moreover, node selection during block generation or model integration is usually fixed or handled without trust based evaluation, which leaves the system open to adversarial behavior. The lack of robust trust and security mechanisms raises the likelihood of corruption in the global model.

Sustaining reliability and active engagement in blockchain driven FL requires incentive mechanisms that can adapt to changing conditions. Many existing designs still depend on basic contribution based rewards, which overlook the complexity of IoT environments. Improved incentive frameworks should consider not only how often or how much a node contributes, but also the value and stability of those contributions, the efficiency of resource usage, communication reliability, and device readiness. In addition, robust trust evaluation is essential to reduce the impact of malicious or underperforming participants. Leveraging machine learning or reinforcement learning for trust assessment can help flag harmful nodes while gradually reinforcing cooperative and beneficial behavior.

Addressing the shortcomings of current designs requiers moving toward smarter, context aware blockchain FL frameworks tailored for IoT. DRL offers a promising pathway by dynamically coordinating operations such as channel distribution, block leader selection, and block size tuning in response to device capabilities and current network status. Pairing layered blockchain structures with hybrid consensus mechanisms can also enhance both scalability and energy performance. Incorporating private or consortium blockchains with edge based aggregators also reduces consensus workload and delays while preserving decentralization and security. Despite its strong potential, blockchain FL integration still demands further exploration to address its open challenges and to create IoT systems that balance privacy, performance, affordability, and long term sustainability.

For safe, flexible, and resource-efficient intelligence in IoAT systems, the suggested Hyper-Learning Framework combines Federated Learning (FL), blockchain, and Deep Reinforcement Learning (DRL) into a single control loop. In this approach, the blockchain layer guarantees integrity, auditability, and incentive management through verifiable update recording, while FL permits privacy-preserving local training.

In response to current network and device conditions, such as energy availability, update quality, and consensus delay, a DRL controller dynamically modifies system parameters like client participation, aggregation intervals, block configuration, and resource allocation. Actions alter FL and blockchain parameters, and rewards represent trade-offs between accuracy, communication overhead, and energy consumption. The interaction is based on a closed-loop decision process. The closed-loop optimization in the proposed Hyper-Learning Framework follows the DRL-driven control process illustrated in Fig. 2, where system states, actions, and rewards jointly guide adaptive federated learning and blockchain operations.

Future research should concentrate on adaptive incentive models appropriate for heterogeneous IoT contexts, scalable FL architectures, and lightweight consensus procedures. Real-world implementation requires enhanced interoperability and standardised assessment. Future studies might investigate DRL-driven incentive models, for instance, in which incentives are dynamically modified according to client dependability, energy contribution, and update quality. High-quality and energy-efficient participants are rewarded more in such a system, whereas malicious or unreliable nodes are gradually penalised. Using DRL to adjust incentive mechanisms across heterogeneous IoT domains—for example, giving priority to energy-aware rewards in wearable healthcare systems and latency-sensitive awards in industrial IoT—is another interesting avenue. These illustrations show how clever incentive design can enhance sustainability, equity, and participation in federated learning enabled by blockchain technology. All things considered, the Hyper-Learning Framework offers a useful path towards obtaining scalable and secure distributed intelligence in IoAT systems.