Research on IDS has evolved across multiple dimensions. The work in [8] presents a broad survey covering application domains, preprocessing techniques, attack detection methods, evaluation metrics, author collaborations, and datasets, offering a detailed landscape of IDS research. Similarly, [9] introduces fundamental IDS concepts and common detection techniques, comparing machine learning- and deep learning-based methods as well as their associated training datasets. Feature engineering for anomaly detection is emphasized in [10], which categorizes IDS according to machine learning, deep learning, and swarm or evolutionary algorithms, and discusses datasets and performance assessment methodologies.

More recent studies shift toward distributed and collaborative approaches. In [11], FL is applied to anomaly detection in IDS, with discussions of privacy protection alongside challenges such as latency, misjudgment, and poisoning attacks. The work in [12] categorizes IoT IDS security issues using a two-tier taxonomy, proposing an integrated platform that combines explainable AI, FL, and social psychology. Likewise, [13] investigates data security and privacy in IoT environments, highlighting blockchain's potential to enhance trust and integrity in machine learning-based IDS. The joint use of FL and blockchain is explicitly analyzed in [14], identifying how the two technologies complement each other in preserving privacy and enabling secure collaboration. Complementary to this, [15] reviews historical privacy and security challenges in IoT and categorizes research efforts applying blockchain and machine learning to safeguard IDS.

A comparative overview of these studies is provided in Table 1. While these surveys contribute valuable insights into IDS, FL, and blockchain, most remain fragmented, addressing isolated aspects without a unified perspective. In particular, systematic analysis of privacy and trust issues in blockchain-federated IDS is still lacking, and no comprehensive taxonomy currently exists. This gap motivates the present work, which consolidates prior contributions and provides a structured view of challenges and perspectives.

Table 1

Comparison of existing surveys on IDS, federated learning, and blockchain.

The key contributions of this paper can be summarized as follows:

We provide a systematic survey of IDS that incorporate FL and blockchain. Their architectures, security objectives, and application scenarios are analyzed, highlighting both strengths and limitations.

The remainder of this paper is organized as follows. Section 2 introduces the background of IDS, FL, and blockchain, and summarizes representative studies in each area. Section 3 analyzes privacy and trust issues in federated IDS, proposes a taxonomy of threats and defenses, and reviews representative studies across domains. Section 4 discusses key challenges of blockchain-federated IDS, including communication efficiency, non-IID data distribution, deployment on resource-constrained IoT devices, incentive mechanisms, and system security, and outlines potential future directions. Finally, Section 5 concludes the paper by highlighting the main findings and prospects for secure and privacy-preserving IDS.

As the Internet evolves and the number of connected devices continues to grow, cyberattacks targeting heterogeneous devices and users have become increasingly complex and diverse. IDS are designed to monitor and analyze activities in networks or hosts, identify potential malicious behaviors, and trigger alerts or interventions. Research on IDS has been active for decades, and this subsection provides an overview of their taxonomy, detection mechanisms, and the advances brought by machine learning (ML) and deep learning (DL).

Blockchain is a decentralized and tamper-resistant public ledger maintained over peer-to-peer networks. It enables multiple participants to securely record, verify, and share transaction data without centralized control. Each node, including those deployed at the edge or on Mobile Edge Computing (MEC) servers, participates equally in validating transactions under a consensus mechanism [24]. In this subsection, we summarize blockchain's key components, taxonomy, and consensus mechanisms, with emphasis on how these features contribute to data security and trust in distributed IDS.

Federated learning, first introduced in [51], has emerged as a promising paradigm for privacy-preserving collaborative training. Traditional centralized machine learning requires collecting all raw data on a central server, which raises significant privacy and security concerns. In contrast, FL distributes model training to edge devices, where participants perform local training on their data and only share model updates with a central aggregator [52]. This avoids direct transmission of sensitive information and substantially reduces the risk of privacy leakage.

Predicting and preventing malicious traffic is the primary purpose of Network Intrusion Detection Systems (NIDS). Machine learning- and deep learning-based IDS heavily depend on large-scale datasets to differentiate benign and malicious traffic. Traditionally, centralized IDS are trained on aggregated datasets and then deployed at endpoints. However, with the rapid growth of IoT and edge computing, data are increasingly generated and stored at distributed nodes. Centralized IDS suffer from scalability issues in model training, malicious behavior detection, and data transmission. Moreover, heterogeneity among devices leads to severe data imbalance, resulting in classification bias, poor generalization, and reduced detection performance. To address these issues, techniques such as re-sampling, synthetic data generation, collaborative learning, and ensemble methods have been applied in distributed IDS [56, 57, 58]. Table 4 summarizes widely used IDS datasets (e.g., KDD99, NSL-KDD, UNSW-NB15, CICIDS2017, CSE-CIC-IDS2018), which continue to serve as benchmarks for evaluating both detection performance and privacy-related risks in IDS research.

Table 4

Overview of commonly used public dataset in intrusion detection.

Beyond imbalance, data privacy and security remain critical concerns. Multi-party data collection and centralized training risk exposing sensitive participant information. Direct aggregation of raw traffic data may violate user privacy and introduce security vulnerabilities, while transmission of unencrypted data creates additional risks of theft and tampering. Studies have shown that IoT traffic often leaks privacy-sensitive information about users, devices, and platforms [59]. Static analysis tools have further revealed that many IoT applications expose sensitive flows during runtime [60]. These risks reduce participants' willingness to contribute data and undermine the credibility of collaborative IDS.

From a classification perspective, sensitive data in distributed IDS can be divided into three categories [61, 62]: (i) input data, such as traffic traces, IP addresses, and user identifiers, which may directly reveal private information; (ii) built-in data, including attack signatures, anomaly profiles, and detection models, which attackers can exploit if exposed; and (iii) generated data, such as detection results, timestamps, and alerts, which may disclose user identities or system behavior when shared across IDS nodes.

During communication, IDS data are particularly vulnerable to man-in-the-middle (MITM) attacks. As illustrated in Figure 4, attackers may intercept or alter traffic without detection, leveraging techniques such as ARP spoofing, DNS spoofing, and SSL hijacking [63]. Such attacks compromise both privacy and trust in distributed detection systems.

Figure 4.

MITM Attack. This figure illustrates how an adversary can intercept or modify traffic between distributed IDS nodes.

In traditional IDS trained on centralized data, data protection relies mainly on encryption and obfuscation. However, transmitting raw data during communication not only increases energy overhead but also poses privacy risks, especially in large-scale distributed environments where participants differ in data structures and operating conditions. FL addresses these challenges by allowing participants to keep their data locally while only sharing model parameters with a central server. This decentralized paradigm reduces the risk of sensitive data leakage and provides stronger privacy guarantees, thereby encouraging broader participation in collaborative intrusion detection.

A typical FL-based IDS workflow involves distributing a global model from a central server, training it locally on participant devices, and then transmitting updated parameters back for aggregation. For example, [68] developed an IoT intrusion detection scheme where participants locally trained a deep learning model and uploaded updates, which were then aggregated using federated averaging (FedAvg). This represents the standard application of FL in IDS scenarios.

Different detection architectures have been deployed under FL-based IDS frameworks. In [69], a CNN-GRU hybrid model was adopted, combining convolutional layers, gated recurrent units, and fully connected layers, trained locally and aggregated globally. In [70], an improved Simple Recurrent Unit (SRU) was introduced to reduce computational cost and mitigate gradient vanishing, demonstrating applicability in Industrial Control System (ICS) networks. Beyond IoT and ICS, FL-based IDS solutions have been extended to diverse domains: a two-stage leakage detection scheme in vehicular CPS (VCPS) [71], adaptive FL algorithms in satellite–terrestrial integrated networks (STIN) [72], FedBatch aggregation in maritime transportation systems [73], and DEW-Cloud-based intrusion detection in medical IoT [74]. These studies highlight the versatility of FL in protecting sensitive data across heterogeneous environments.

Overall, IDS built on FL significantly increase the difficulty for attackers to access raw data. Data owners maintain greater control and ownership, deciding when and how to share updates. Since only parameters are exchanged, FL enables cooperative learning under strict data regulations, while avoiding direct exposure of private datasets. However, limitations remain: model updates may still leak information through inference, aggregation servers may pose single points of failure, and communication overhead can be non-trivial in large-scale deployments. Table 6 summarizes the advantages and limitations of FL-based IDS.

Table 6

Advantages and limitations of FL-based IDS.

Research [75, 76, 77] has shown that gradient information may leak sensitive local data, making FL face both credibility and privacy risks. A major challenge is that participants must trust a central server for model aggregation. Even under a semi-trusted assumption, the server may act honestly but remain curious, attempting to infer private information from updates. To mitigate such risks, several privacy-preserving techniques have been explored.

Differential Privacy (DP). DP [78] introduces statistical noise into training data or model updates, reducing the risk of private information extraction. In FL, participants train locally and add Gaussian or Laplace noise before uploading parameters, thereby protecting privacy while retaining model utility. Work in [79] evaluated the impact of DP on intrusion detection in IoT systems, showing that aggregation functions such as Fed+ can preserve model accuracy within acceptable ranges despite the added noise.

Homomorphic Encryption (HE). HE [80] enables secure computation over encrypted data without requiring decryption. In [81], the Privacy-Enhanced Federated Learning (PEFL) framework employed linear homomorphic encryption, where participants add noise to local gradients before transmission. Similarly, [82] combined ϵ-DP with Paillier homomorphic encryption to secure parameter updates in fog computing scenarios, preventing even honest-but-curious servers from learning sensitive information.

Secure Transmission Protocols. During communication, model updates remain vulnerable to man-in-the-middle attacks. To address this, [69] designed a secure protocol based on Paillier HE, supporting key generation, encryption, aggregation, and decryption. Likewise, [83] applied secure key exchange protocols in an intelligent factory setting, employing RING-LWE shared keys with AES ciphers to protect communication channels against tampering.

Table 7 summarizes representative approaches for privacy-preserving federated learning, classified by the stage at which data are protected (in storage vs. in transmission).

Table 7

Research on data privacy–preserving approaches in federated learning.

While cryptographic techniques enhance privacy during storage and transmission, they do not address the systemic limitations of centralized aggregation. In particular, reliance on a single aggregation server reduces system robustness and introduces a single point of failure. To overcome these drawbacks, blockchain has been integrated into FL frameworks.

As a decentralized, immutable, and traceable ledger, blockchain supports distributed model aggregation without a central server. In blockchain-based FL, participants upload local model updates to designated miner nodes. Miners exchange, verify, and reach consensus on updates, after which a new block containing aggregated results is appended to the ledger. Participants then download the global model from the blockchain, as illustrated in Figure 5. This process not only eliminates single points of failure but also strengthens auditability and participant incentives.

Figure 5.

FL with integrated blockchain.

Building on the general integration of blockchain and FL discussed in Section 3.3, recent studies have explored concrete applications of these techniques in diverse network environments. The overall architecture of such an integrated NIDS is illustrated in Figure 6, which depicts the key interactions among edge devices, blockchain consensus mechanisms, and global model distribution for secure intrusion detection.

Vehicular Networks (V2X). IDS in vehicular networks face challenges due to the limited computational resources of vehicles and roadside units (RSUs), which act as edge nodes. Traditional federated learning-based IDS may suffer from security concerns in data storage and sharing. Research [87] proposed a blockchain-enhanced federated learning IDS where RSUs serve as mining nodes to collect, train, and aggregate model parameters from vehicles. An accuracy-proof consensus protocol ensures that RSUs providing higher-accuracy models write updates to the blockchain. A trust-based incentive mechanism further strengthens reliability, while dynamic secret sharing is used to protect uploaded gradients. On the KDDCup99 dataset, the system achieved 94% accuracy, demonstrating the effectiveness of this approach.

Drone Networks. Drone devices typically operate under constrained resources, making them vulnerable to intrusion and yielding small, imbalanced datasets. To address these challenges, [88] proposed a collaborative IDS based on FL and blockchain, integrating a Conditional GAN (CGAN) and LSTM architecture. Multiple mobile edge computing (MEC) servers form a blockchain network to store global models and achieve consensus. Privacy is preserved by adding Gaussian noise to parameter transmissions. Evaluated on CIC-IDS2017, the FL-CGAN-LSTM model achieved accuracy above 99% for large classes (e.g., Normal, DDoS, PortScan) and improved small-class detection (e.g., Infiltration, Bot) from 29% to over 95% after dataset extension.

Industrial Edge of Things (IEoT). In heterogeneous industrial IoT environments, devices run under diverse systems and generate privacy-sensitive data. To secure such systems, [89] proposed Fed-Trust, a blockchain-based FL intrusion detection model employing a Time Convolutional GAN (TCGAN) for semi-supervised learning on partially labeled data. A blockchain-based reputation system records gradients and ensures credibility, while fog computing offloads mining to enhance scalability. The use of Residual Temporal Convolution (RTC) and Causal Dilated Convolution (CDC) layers enables effective modeling of multivariate time series data. Credit Contracts provide lightweight incentive and trust mechanisms, improving throughput compared to Bitcoin-like blockchains.

Internet of Medical Things (IoMT). Medical IoT environments involve highly sensitive data and demand strict privacy guarantees. To address this, [90] introduced a blockchain-based FL architecture for intrusion detection in IoMT. The framework eliminates the central server, instead using IoT gateways as participants in a federated network, while Hyperledger Fabric maintains distributed ledgers. Local updates and aggregated global models are recorded as immutable transactions, complete with timestamps for auditability. This design enhances trustworthiness and makes tampering with medical intrusion detection data significantly more difficult.

Metaverse-Oriented IDS. The metaverse relies on extensive IoT connectivity and thus inherits traditional IoT security threats. To address this, [91] proposed MetaCIDS, a collaborative IDS combining blockchain and FL. It uses an unsupervised autoencoder for zero-day attacks and a supervised classifier for known attacks, with model updates managed by blockchain using practical Byzantine fault tolerance (PBFT). Intrusion alerts are validated by smart contracts before being broadcast as global alerts. Participants are incentivized through rewards or penalties depending on their detection accuracy. On datasets such as UNSW-NB15, CSE-CIC-IDS2018, and CIC-IDS2017, MetaCIDS achieved 95–99% accuracy, outperforming classical ML models like SVM, KNN, and LightGBM.

Figure 6.

NIDS architecture integrating blockchain and federated learning.

In FL and blockchain-based systems, frequent data exchanges among distributed devices may cause network congestion, especially when the number of participants increases significantly. Such congestion results in higher communication latency and potential data loss, which can become a bottleneck affecting overall system performance. The large volume of parameter transmissions and the delays introduced by multi-party exchanges are thus key factors influencing communication efficiency.

Reducing the communication overhead is a potential solution. In [92], a universal vectorization method is proposed, where each participant divides model parameters into multiple sub-vectors and independently quantizes them. The quantized sub-vectors are then aggregated at the server to form a global codeword. In subsequent rounds, participants match their sub-vectors with the codeword to generate indices, thereby reducing the amount of transmitted data.

Communication efficiency can also be improved by adjusting federated communication strategies and blockchain computation mechanisms. For instance, [93] introduces a blockchain-based FL framework for privacy-aware vehicular communication networks. The study evaluates parameters such as retransmission constraints, block size, block arrival rate, and frame size, and examines their impact on performance. Results show that as the signal-to-noise ratio increases, system latency decreases. However, channel fading in wireless communication can lead to data transmission failures, which increase blockchain fork probability and thus overall latency. Moreover, in cases of wireless loss, the latency observed by miners and vehicles is higher than that measured at the network layer. Variations in the number of vehicles and miners predictably affect learning latency, and adjusting block arrival rates can help achieve dynamic optimization. Determining the optimal arrival rate requires jointly considering the number of miners and channel propagation delay.

In [94], communication efficiency is enhanced by reducing the parameter size transmitted to the aggregation server and by optimizing transmission resource allocation. The approach employs asynchronous model updates, where participants update at different frequencies based on assigned weights that reflect their local model update frequency. This decentralized updating strategy lowers transmission overhead and avoids the synchronization delay of waiting for all users to complete training, thereby reducing overall execution time. Furthermore, the system leverages state-based digital twin technology to monitor device status and dynamically allocate communication resources according to device conditions and computing capabilities, further reducing communication time and energy consumption.

Research [95] has shown that neural networks trained with the Federated Averaging algorithm suffer a significant accuracy decline when dealing with non-independently and identically distributed (non-IID) data, primarily due to weight divergence. In federated IDS, the heterogeneous operating environments of participating devices lead to non-uniform training data, with substantial variations in distribution across devices. Such non-IID characteristics can result in model divergence during training and aggregation, particularly in horizontal FL settings.

To mitigate these challenges, [95] introduces an improved Federated Averaging approach based on a data-sharing strategy. In this method, a portion of uniformly distributed global data is provided by the aggregation server to participant devices at the initial stages of training, ensuring greater consistency across datasets. Additionally, pre-training the global model on the shared data before distributing it to participants enhances model accuracy compared to initializing with random weights.

Other strategies have also been proposed. In [96], a hierarchical clustering process measures the similarity of participants' local updates during training and iteratively merges the most similar clusters, allowing each cluster to train a dedicated model. In [97], a sparse ternary compression (STC) protocol is designed to improve accuracy on non-IID data while reducing communication overhead. By introducing sparsity and ternarizing model parameters, the protocol effectively extracts features while lowering transmission costs. Furthermore, [98] systematically categorizes solutions to non-IID issues into three groups: (i) data-based approaches, such as data sharing and augmentation, which improve training distribution; (ii) algorithm-based approaches, including participant weighting, meta-learning, and regularization strategies; and (iii) system-based approaches, such as clustering participants to build cluster-specific models.

For example, in a smart transportation deployment [99], roadside units (RSUs) at urban intersections and suburban highways collect traffic with markedly different class priors and temporal patterns, yielding non-IID local datasets. A practical avenue is to combine clustered federated learning with domain adaptation: RSUs are first grouped by similarity of update statistics, each cluster trains a dedicated model via hierarchical aggregation, and lightweight adaptation layers align feature distributions when models are transferred across clusters. This design can mitigate weight divergence, improve convergence speed, and raise detection accuracy under non-IID conditions while keeping communication overhead comparable to standard FL.

IoT is one of the primary environments for deploying IDS. While blockchain ensures data privacy and credibility, it was originally designed for powerful computing platforms rather than resource-constrained IoT devices. In blockchain- and federated-learning-based IDS, workload heterogeneity and device capability disparities significantly impact model training performance. Many IoT edge devices with limited computational resources introduce delays in block validation, storage, and consensus during training.

The study [100] evaluated two parameter update processes in blockchain–federated learning models. When the parameter server directly received and processed weight updates from deep learning servers, an additional computational overhead of 5%–15% was observed. Conversely, when updates were validated and stored on the blockchain, delays from block generation and confirmation became more pronounced. To address these issues, [101] proposed a deep reinforcement learning–based resource allocation framework. Treating blockchain transaction arrivals as a Poisson process, the system was modeled as an M/M/c queue with additional propagation delays. Optimization problems concerning device resources, block generation rates, and mining operations were formulated within a reinforcement learning framework, with Deep Q-Networks (DQN) used to derive optimal allocation strategies. Similarly, [102] introduced a digital twin network model integrating blockchain and FL into edge environments. By synchronizing with physical IoT systems, the digital twin network enabled real-time analysis and optimization of device operations, reducing resource disparities and overall system costs.

Due to computationally intensive mechanisms such as proof-of-work and encryption, blockchain deployment imposes high demands on computation, storage, and bandwidth. This creates significant pressure when applied to IoT devices. Designing lightweight blockchain platforms and consensus mechanisms tailored for IoT has thus become an important research direction. For example, [103] proposed a collaborative multi-proof-of-work consensus mechanism using a Collaborative Index (CI) framework to dynamically adjust mining difficulty, thereby reducing overhead. [104] developed a lightweight blockchain platform by improving Hyperledger Fabric and Mysitko, constructing a distributed microservices architecture with Golang and Scala to enhance concurrency and throughput. [105] introduced DV2G, a lightweight blockchain protocol employing a game-theoretic transaction model and a directed acyclic graph (DAG) structure to enable parallel transaction processing. Finally, [106] proposed the PoBT consensus protocol, which increases transaction rates in IoT systems by reducing block creation and verification costs, while employing local transaction isolation to lower storage and bandwidth requirements.

Participants providing high-quality data and models are crucial to improving system performance in FL and blockchain-based IDS. Designing appropriate incentive mechanisms ensures that such participants are sufficiently rewarded, compensating for their resource consumption and encouraging sustained contributions. Efficient incentive strategies not only improve model quality but also attract more participants, thereby expanding data coverage and enhancing detection accuracy.

Several incentive mechanisms have been proposed in the literature. In federated learning– and blockchain-based vehicular networks, [87] proposed a trust-based reward mechanism that links model accuracy with trust evaluation. When a participant's model achieves higher accuracy than others, its trust score increases; conversely, lower accuracy reduces the score. Participants can then query trust values recorded on the blockchain before verifying the received models. Models from participants whose trust value falls below a threshold are considered unreliable, even if their accuracy is high, thereby deterring malicious or inconsistent contributions.

Economic approaches have also been explored. The study [107] introduced an incentive mechanism grounded in contest theory from auction theory. By modeling repeated competitions among rational participants, the mechanism ensures protocol compliance and profit maximization. A reward policy further encourages workers to follow the protocol while receiving proportional benefits, with incentive compatibility theoretically proven.

Value-driven incentives have been proposed as well. In [108], participants contribute local data for collaborative training by initiating blockchain transactions and paying transaction fees. Fees are adjusted based on data volume—participants with larger datasets typically pay lower fees—while negotiation of total fees is supported. Rewards are granted to participants who successfully generate new blocks, aligning incentives with system goals.

Finally, privacy-aware incentive mechanisms have been investigated. The work [109] integrates privacy protection into the incentive process by delegating verification tasks to other participants. This design reduces the computation, communication, and storage burden on individual nodes, while also preventing adversaries from tracking users via IP addresses.

Although blockchain and FL enhance data privacy and security in IDS, such systems remain susceptible to adversarial and poisoning attacks.

A subset of participants may behave maliciously by uploading carefully crafted poisoned parameters to the server, contaminating the global model, as illustrated in Figure 7. Models generated from these parameters typically retain good performance on most normal samples while deliberately misclassifying specific malicious traffic targeted by the attacker. To enhance the effectiveness of poisoning, [110] proposed a toxic data generation algorithm, Data Gen, based on Generative Adversarial Networks (GAN). In this approach, the central server shares global model parameters with the GAN discriminator, while the generator produces poisoned samples. Since direct uploads of poisoned parameters may reduce their effectiveness during model averaging, the method introduces a scaling factor to amplify gradient changes, ensuring the persistence of poisoning effects.

Figure 7.

Poisoning attack in federated training.