Secure IIoT architecture with blockchain-enabled anomaly detection for water distribution cyber–physical systems Authors: Abdu Salam, Ahmed Thair Shakir, Qaisar Javaid, Barbara M Masini, Masood Ahmad, Ishtiaq Wahid Published: 2026 ABSTRACT AbstractIndustrial Internet of Things (IIoT)-based technologies are becoming increasingly popular in water-distribution cyber-physical systems (CPS) for automated, real-time control and information-based decision-making. Nevertheless, this increased digital connectivity has also expanded the attack surface, allowing adversaries to interfere with sensor data and control logic, and to cause harmful hydraulic disturbances. To overcome them, this paper proposes a comprehensive secure IIoT architecture that integrates blockchain-based data integrity with a sophisticated anomaly-detection framework tailored to a water-distribution CPS. The architecture uses a permissioned blockchain to provide tamper-resistant recording of sensor telemetry, device identity validation via smart contracts, and verifiable provenance for the operator and forensic analysis. To supplement the trust layer, a hybrid anomaly-detection model based on LSTM autoencoders with cyber–physical feature fusion can provide real-time detection of both sudden and covert attacks. Three popular benchmark datasets, BATADAL, WADI, and SWaT, are used to evaluate the proposed system, simulated hydraulic networks, and full-scale industrial testbeds. The results of the experiments show that the proposed framework is more accurate, has higher recall, and achieves a higher AUC than classical machine-learning and modern deep-learning baselines. The cross-dataset analysis demonstrates the framework’s ability to be transferred into a cross-dataset and heterogeneous water and industrial CPS environment, where performance depends on the density of sensors, complexity of processes, and distribution of attacks. The result shows the strength of the proposed model and the potential necessity of further domain adaptation-based methods in the case of highly heterogeneous deployments.Keywords: Industrial Internet of Things (IIoT), Cyber–Physical Systems, Water Distribution Networks, Blockchain Security, Smart Contracts, Anomaly Detection, Deep LearningSubject terms: Energy science and technology, Engineering, Mathematics and computing Introduction The water distribution system has been digitized over the last several years, and utilities have begun implementing modern monitoring and control systems to enhance operational efficiency1. They now form part of cyber-physical systems (CPS), which include embedded controllers, communication links, and field devices that interact in real time with hydraulic processes. IIoT technologies can capture fine-grained measurements of flow, pressure, and quality parameters, enabling utilities to build better situational awareness and automate multifaceted operational functions. Nevertheless, heightened connectivity has also expanded the attack surface, making water infrastructure vulnerable to cyber intrusions that can cause massive physical impacts2. Today’s utilities employ networked sensing and automated control of distribution pipelines, reservoirs, and pumping plants. This has enhanced reliability and reduced operational costs at the expense of excessive dependence on computers, which are liable to failure in water systems. It intensifies the need to create architectures that enhance data integrity, ensure secure communication, and enable timely detection of anomalies before they lead to failure, as the number of attacks on water facilities, including unauthorized access, sensor tampering, and operational interference, grows3. These trends lead to a full-fledged system that incorporates safe IIoT connectivity, powerful data processing, and intelligent anomaly detection. CPS Water distribution has several problems that distinguish it from conventional IT networks. The old SCADA protocols lack encryption or authentication, making them easy to compromise by unscrupulous individuals4. The heterogeneous IIoT devices also complicate security management because they have varied firmware, inconsistent security features, and varying update schedules. The threats to such systems are not usually divided into digital and physical effects. For example, injection attacks involving false data can cause controllers to perform harmful physical actions, e.g., pump speed or tank setpoint adjustments5. The devices of IIoT lie at the core of the collection and transmission of operational information; however, because they are widely used, they are vulnerable to tampering and spoofing. One solution may be blockchain, an immutable, decentralized registry that can store sensor data and enforce trusted interactions between distributed components. As shown in the literature, blockchain can be used to safeguard device identities, record state changes, and enable traceable audit functions within a CPS environment6,7. Smart contracts can also be used to validate rules, allowing a programmable level of security. Together with IIoT, blockchain will enhance data provenance, minimize the risk of manipulation, and support forensic investigation in the aftermath of a cyber-attack, which would be particularly useful for critical infrastructure such as water systems8,9. Traditional anomaly detection in modern water systems is inadequate because hydraulic operations are highly dynamic and intricate. A large number of cyber-physical attacks are deliberately low-profile, aiming to avoid threshold-based detection by emulating natural variation. The capabilities provided by machine learning and deep learning are superior because they capture cross-sensor correlations, temporal dependencies, and complex operational patterns10. Recent research has demonstrated that deep-learning techniques have become increasingly useful for anomaly detection in IIoT and cyberphysical systems, owing to their ability to capture nonlinear temporal effects and multivariate operational patterns11. Simultaneously, blockchain-based systems have been explored to enhance data integrity, trust management, and secure data processing in distributed industrial settings. These two directions are further explained by the studies recommended by the reviewer, which focus on intelligent anomaly detection and blockchain-assisted secure data processing in IIoT environments12. Recent research also demonstrates that intelligent computation is able to aid in the management and optimization of water infrastructure. Marar and Marar developed an efficient algorithmic water delivery system and smart water metering in water scarcity areas, which shows the importance of algorithmic monitoring in improving operational efficiency in water systems13. In another, but methodologically similar, engineering field, Liu et al.14 applied a machine learning-based rapid design approach for hydrodynamic shape optimization of underwater vehicle pedigrees, demonstrating the potential of data-driven models to speed up the design of complex physical systems. The studies show the growing significance of intelligent algorithms in the fields of infrastructure and fluid-related cyber-physical systems. They don’t directly solve the problem of secure IIoT telemetry provenance, integrity verification using a blockchain, and the detection of anomalies in an adversarial water distribution CPS. In the present study, we build upon the aforementioned work and study security-oriented cyber-physical monitoring, where data authenticity, attack detection, and forensic auditability are all required. This work proposes a single security architecture for water-distribution CPS that combines IIoT sensing, blockchain-based data integrity, and the most sophisticated anomaly-detection techniques. The key contributions are the following: A unified secure IIoT architecture designed explicitly for water-distribution CPS, enhancing device trust, data authenticity, and communication security. A unified secure IIoT architecture designed explicitly for water-distribution CPS, enhancing device trust, data authenticity, and communication security. A blockchain-enabled integrity framework that provides tamper-resistant logging of sensor data and anomaly events, supporting transparent auditing and forensic analysis. A blockchain-enabled integrity framework that provides tamper-resistant logging of sensor data and anomaly events, supporting transparent auditing and forensic analysis. A hybrid anomaly-detection model that fuses physical sensor data with cyber-layer features to identify subtle, multi-stage, and stealthy attacks more effectively than traditional methods. A hybrid anomaly-detection model that fuses physical sensor data with cyber-layer features to identify subtle, multi-stage, and stealthy attacks more effectively than traditional methods. Comprehensive cross-dataset validation using BATADAL, WADI, and SWaT to demonstrate robustness, generalizability, and operational feasibility in diverse water-infrastructure environments. Comprehensive cross-dataset validation using BATADAL, WADI, and SWaT to demonstrate robustness, generalizability, and operational feasibility in diverse water-infrastructure environments. The paper is structured as follows. Section 2 reviews related work, and Sect. 3 presents the system model and threat assumptions. Section 4 presents the proposed secure IIoT architecture, a blockchain-based anomaly-detection system. The results are discussed in Sect. 5. Section 6 concludes the paper. Related work The increased interconnection of critical infrastructure has driven rapid research on IIoT security in industrial CPS. Initial research found that, in the vast majority of cases, operational technology (OT) rooms were not designed to be connected to the Internet, and even older communication protocols, such as Modbus and DNP3, were vulnerable to spoofing, replay attacks, and unauthorized access15,16. With IIoT sensors and gateways being implemented across industrial sectors, the attack surface has increased manifold; compromised field devices have been shown to cause cascading physical failures, especially in water, energy, and manufacturing systems17,18. Recent work indicates that IIoT architectures require a secure architecture with device authentication, end-to-end encryption, and distributed trust. Some of the methods are hardware-based identity modules, secure firmware signing, and zero-trust segmentation of industrial networks19,20. Also, a few papers report that conventional perimeter security is inadequate for CPS and suggest in-network sensor technologies that combine physical and computer telemetry to identify early intrusions21,22. Although this has been advanced, it doesn’t remain easy to achieve heterogeneous IIoT ecosystems, as not all field devices have sufficient processing capabilities, and security compliance is not a universal concern among vendors. Additional technologies have become interesting additions to CPS security, with blockchain and distributed ledger technologies (DLT) enabling tamper-resistant logging, decentralized verification, and trustless coordination. Solutions for securing sensor data, implementing device identity, and validating control commands in the industrial sector have been proposed within a blockchain framework23,24. Immutability is applied in these systems to prevent undetected manipulation of operational records, which supports forensic analysis and auditability of critical infrastructure. Blockchain has been used for supply-chain tracking, smart energy, and distributed control in the industrial IoT, where its benefits include a lack of centralized authority and increased robustness against single-point failures25,26. Lightweight consensus algorithms, such as PBFT, PoA, and Directed Acyclic Graph (DAG)- based ledger consistency, have also been used to meet CPS latency requirements. In contrast, more resource-intensive protocols, such as PoW, have been found unsatisfactory27,28. However, blockchain has yet to gain popularity in water distribution systems. Recent work is the first to consider blockchain-based digital twins, safe telemetry pipelines, and decentralized anomaly-checking systems9,29. The recent literature has examined the security, trust, privacy, and resilience implementations in the IoT-enabled systems, such as lightweight cryptography, blockchain-based secure networking, intelligent water-management schemes, and privacy-preserving data sharing30,31. These trends inspire our research and aim specifically at water-distribution cyberphysical systems through a hybrid approach to anomaly detection and permissioned blockchain-based integrity assurance of safe surveillance and tamper-evident event reporting32. Traditional practices rely on rule-based alarms or statistical thresholds, and they are unable to detect subtle or coordinated attacks. Standard datasets, such as BATADAL, WADI, and SWaT, have spurred more serious assessment of detection techniques33. More recent research discusses data-driven methods that incorporate signals from physical sensors (pressure, flow, tank levels) with network traffic signals to identify anomalies caused by cyber intrusions, equipment failures, or misuse of the equipment34. The methods include probabilistic modeling, clustering, and multivariate temporal analysis. A hybrid intrusion detection system that runs concurrently at the network and process levels has demonstrated potentially useful performance, minimizing false positives and being more sensitive to stealthy false-data injection attacks35,36. Nevertheless, it is frequently dataset-dependent, and generalizable models tested across a variety of water-network environments are required. The intrusion detection, fault diagnosis, and predictive maintenance of industrial CPS have been extensively implemented using machine learning (ML) and deep learning (DL). SVMs, random forests, and PCA are classical methods that enable lightweight anomaly detection, which is why they are compatible with edge computing assets37. In the meantime, deep architectures, such as LSTM networks, autoencoders, CNN-LSTM hybrids, and graph neural networks, learn temporal relationships and spatial correlations common in CPS operations38,39. As recent studies have highlighted, it is crucial to integrate security, privacy protection, and smart decision-making in contemporary cyber-physical and IoT contexts40. Previous studies have also established that hardware-aided authentication and blockchain-based protection can enhance resilience against advanced AI-driven attacks in interconnected systems41. Simultaneously, reinforcement-learning-based optimization has shown strong potential for adaptive resource management and path control in distributed edge-assisted networks42. Applications in water systems include modeling hydraulic patterns, identifying sensor tampering, and detecting multi-stage sequences of cyber-physical attacks. Research indicates that the reconstruction errors of autoencoders, graph-based reasoning, and spatiotemporal learning achieve much higher performance than fixed thresholds on the WADI and SWaT datasets43,44. However, a significant portion of DL-based solutions has certain limitations, including high computational requirements, limited interpretability, limited cross-dataset generalizability, and an inability to integrate with security measures such as blockchain or secure IIoT communication systems. Table 1 shows the comparative overview of the available research. Table 1Comparative Summary of Prior Studies.Ref.Focus areaMethod/ApproachDataset usedKey limitations 17 IIoT security for CPSSecure gateway + encryptionSynthetic CPSLimited to small networks 18 Legacy SCADA securityProtocol hardeningReal plant logsNo anomaly detection 19 IoT trust and identityDevice authenticationTestbedHigh overhead 45 Blockchain for IIoTDLT trust modelSimulationLatency concerns 25 Blockchain for CPSSmart contractsManufacturingNot water-specific 9 Blockchain + IDSSecure loggingSWaTNo physical-layer fusion 33 Water-network anomaly detectionStatistical modelsBATADALHigh false positives 35 Hybrid cyber–physical IDSMultilayer fusionWADILimited generalization 46 DL for CPS anomaliesLSTM/AE modelsSWaTHigh computational load 43 Water CPS DL detectionGNN-basedWADI/SWaTRequires large datasets Comparative Summary of Prior Studies. The existing studies, as presented in Table 1, typically focus on either secure data exchange, logging through blockchain, or anomaly detection as a design element. For water distribution CPS, however, it is necessary to have a unified architecture that also takes into account the trust of devices, the provenance of the telemetry, the scoring of anomalies in the cyber/physical space, and the deployable integration of the gateway. For the sake of better understanding the novelty of the proposed framework, Table 2 compares the proposed architecture with representative previous CPS security solutions at a component level. The proposed approach differs from previous approaches, which focused primarily on either IDS accuracy or data integrity as their primary metrics, by introducing a hybrid framework that integrates permissioned blockchain-based provenance, smart contract-based device and transaction validation, LSTM autoencoder-based temporal anomaly detection, cyber-physical score fusion, and cross-dataset validation across BATADAL, WADI, and SWaT. Table 2Component-level novelty comparison between prior CPS security frameworks and the proposed framework.Study categoryTrust modelDetection pipelineDeployment layerBlockchain roleCyber-physical fusionCross-dataset validationMain limitation addressed by the proposed workSecure gateway-based CPS studiesCentralized gateway trustRule-based or conventional IDSGateway or SCADA layerNot includedLimitedUsually not includedLacks tamper-evident provenance and distributed verificationBlockchain-based IIoT studiesDistributed ledger trustUsually not detection-focusedCloud or edge ledgerData logging or access controlLimitedNot usually evaluatedDoes not jointly model hydraulic process anomaliesBlockchain plus IDS studiesLedger-assisted trustIDS with secured loggingEdge or cloudEvent recordingPartialUsually, a single datasetLimited explanation of physical and cyber feature interactionDeep learning-based water CPS IDSModel-based trustLSTM, autoencoder, CNN, GNN, or hybrid detectionAnalytics layerNot includedSometimes includedLimitedDetection results are not linked to immutable provenanceProposed frameworkPermissioned blockchain plus authenticated gateway trustLSTM autoencoder reconstruction plus cyber inconsistency scoringLegacy compatible edge gateway, SCADA, and blockchain validator layerDevice validation, telemetry hash anchoring, anomaly event provenance, and forensic auditExplicit weighted fusion of physical reconstruction error and cyber inconsistency scoreBATADAL, WADI, and SWaTProvides unified detection, provenance, auditability, and deployment feasibility for water distribution CPS Component-level novelty comparison between prior CPS security frameworks and the proposed framework. System model and problem formulation Water distribution CPS architecture The physical hydraulic infrastructure for water distribution is integrated with sensing, actuation, communication networks, and supervisory control to form a water distribution cyber-physical system (CPS). The physical aspect of the system comprises reservoirs, pumps, pipelines, valves, and storage tanks that control flows and pressures within the network. At the cyber layer, real-time measurements of flow rate, pressure, tank levels, pump status, and water quality indicators are obtained from IIoT sensors. These field units send telemetry to programmable logic controllers (PLCs), remote terminal units (RTUs), and a SCADA server, which execute control logic and provide operator visualization. Information exchange occurs via vendor-compliant wired industrial protocols and wireless IIoT networks, often via gateways between the operational technology (OT) layer and the cloud-based environment. Edge or control center data analytics modules process sensor streams to detect anomalies, optimize operations, and predict events. Figure 1 presents the layered architecture of the proposed cyber-physical system of secure water distribution. It also depicts the end-to-end connection of the real water network of the SCADA/PLC and RTU control infrastructure, the communication middleware, routers, and IoT gateways, blockchain and smart-contract services, anomaly-detection engines, and cloud/digital-twin analytics. It also indicates the point of cyberattack entry with the help of the layer of communication. It demonstrates how the proposed framework integrates monitoring, detection of anomalies, and tamper-evident blockchain-based logging on both cyber and physical layers. Architecture of a Water Distribution Cyber–Physical System. Threat model and attack vectors The threat model assumes that adversaries can access or manipulate the cyber or physical layer, or both simultaneously. The attackers seek to compromise IIoT gateways, use insecure protocols, hack sensors, or inject malicious commands, which affect physical processes. According to the existing literature, the most commonly used attacks are false-data injection, spoofing, replay, configuration, denial-of-service (DoS), actuator override, and malicious firmware modification. Due to the interconnection between the water distribution process and other processes, even slight measurement discrepancies can lead to significant hydraulic imbalances. Cyber intrusions may remain stealthy for long periods, particularly if attackers design changes to resemble normal fluctuations. A summary of typical attack surfaces in water distribution CPS is provided in Table 3, categorized by affected component, attack method, and potential physical impact. Table 3Typical attack surfaces in water distribution CPS.ComponentAttack vectorDescriptionPotential impactIIoT SensorsFalse-data injectionTampering with pressure/flow readingsMisleading SCADA decisionsPLC/RTUCommand spoofingUnauthorized actuator controlPump damage, tank overflowGatewaysMan-in-the-middleInterception/modification of dataLoss of integrity & trustCommunication LinksReplay attacksRe-sending valid packetsErroneous system statesSCADA ServerMalware/RansomwareSystem takeover or shutdownOperational paralysisActuatorsForced overrideDirect manipulationHydraulic instabilityCloud AnalyticsData poisoningCorrupting ML/DL trainingFaulty anomaly detectionUser InterfacesCredential theftUnauthorized operator accessConfiguration abuse Typical attack surfaces in water distribution CPS. Problem definition Suppose the water-distribution cyber-physical system (CPS) is observed at discrete time steps . At a given time step, the physical process is modeled by a multivariate sensor-state vector. where is the measured value of the process variable number i-th (i.e., flow rate, pressure, tank level, water-quality indicator, or device-status signal), and n is the number of sensed variables. This notation has been used throughout the manuscript. The discrete time index is represented as t, the physical sensor vector as , the actuator (or control) vector as , the feature vector for the cyber layer as and the length of the temporal observation window as . The nonlinear transition function for the water CPS is unknown, and is represented by the unknown function . The process and measurement uncertainty is denoted by , which is considered to be limited under normal operating conditions. The attack or fault-induced perturbation is denoted by . It is not assumed that all hydraulic variables in the model are strictly stationary; instead, the temporal dependency between variables is assumed to be known from a training window of attacks when there are no attacks, and calibrated when attacks are present through a validation set. The actuator vector is used to represent the control input into the system: where represents the j-th control action, e.g., pump command, valve position, or other supervisory control signal, and m is the number of control channels. In the normal operation conditions, the CPS develops based on the nonlinear state-transition model: with the unspecified nonlinear systems dynamics that control the time dependence of the hydraulic and operational process, and is the collective uncertainty, process, sensor, and measurement noise at time t. The process under observation may fail to respond normally when a network attack, a process malfunction, or a carefully orchestrated cyber-physical disruption occurs. We have deviated by: where is the attacked or perturbed observation and is an additive perturbation value of malicious manipulation, spoofing, communication corruption, actuator-induced deviation, or fault effects. The perturbation can affect one or more sensing or control channels. Stealthy attacks are the hardest to detect since can be small relative to normal operating variations over short-term periods, but over time, it can build up and slowly drive the system to unsafe operating conditions. The objective of the anomaly-detection layer is to learn a decision function: where w is the temporal window length, and are the sequence of sensor states and control inputs over the last w time steps, is the sequence of cyber-layer features, and represents an anomalous state, and represents a normal state. The cyber layer feature vector may include packet rates, command frequencies, inter-arrival delay statistics, retransmission rates, device status changes, actuator command transitions, or other communication side consistency indicators. Based on this, a requested detector must meet when the system operates correctly and when there are cyber-attacks, process errors, or coordinated cyber-physical abnormalities. Meanwhile, the detector must have a minimum number of false alarms during normal system operation. This can be stated as: which is equivalent to reducing the risk of pronouncing an anomaly in case the actual state of the system is normal. In order to facilitate data-integrity checking, every telemetry item accepted is linked to a cryptographic digest: and is a hash collision-free function, the concatenation operator, and is the resulting digest. This digest is pegged to the verified blockchain registry to offer provable provenance to the operational data and events of interest. In order to avoid the duplication of effort, the above formulation is used as the only system-level mathematical model in this study. The later anomaly detection part only translates this model into the definitions of the reconstruction error, the cyber inconsistency score, the fusion rule, the thresholding process, and the anchoring process of the blockchain. Design goals for secure IIoT water distribution The proposed architecture aims at meeting four realistic security and operational objectives for water-distribution CPS. In the first place, it is necessary to maintain the integrity of data and provenance between field sensing and supervisory analytics to verify sensor readings, control messages, and anomalies detected in the future. This necessity prompts the application of a permissioned blockchain in order to pin cryptographic digests of operational data and security events. Second, the system should facilitate concomitant cyber-physical anomalies detection. Process faults, sensor spoofing, command manipulation, and coordinated multi-stage attacks may cause abnormal behavior in water CPS. As such, the physical-process variables and cyber-layer indicators should be merged as the detection layer in order to minimize false alarms in the presence of dynamic hydraulic conditions and enhance sensitivity to stealthy attacks. Third, the architecture should be able to provide safe communication and the trust of the devices. The communication should be performed by field devices, gateways, SCADA servers, and analytics services via authenticated and encrypted channels, and untrusted or compromised nodes should be separated to restrict further lateral movement and decrease the consequences of partial network impact. Fourth, the framework should be capable of being deployed in legacy operational environments. Because most utilities still use installed SCADA servers, PLCs, and protocol stacks with insufficient built-in security, the security controls announced in this paper should introduce limited computational and communication overhead, be configurable via gateway and protocol adaptation, and need not replace the entire infrastructure. Proposed secure IIoT architecture Architectural overview The proposed secure IIoT architecture provides end-to-end security, reliable data provenance, and robust anomaly detection for water-distribution CPS. It is organized into coordinated layers that span field-level sensing, communication infrastructure, blockchain-based trust, and intelligent data analytics. The architecture is modular, so integrating small modules can be done without redesigning the SCADA or PLC systems used by operators. Here, the blockchain layer runs in parallel with prime control to ensure that tamper resistance and verification do not introduce latency in time-sensitive hydraulic processes. The benefits of such a strategy include safe grounds for identifying anomalies, auditing, and further development of digital twin technologies and innovative autonomous water management services. An actual implementation of the proposed framework does not require the replacement of current SCADA servers, PLCs, remote terminal units, and field sensors. Most water-distribution settings have devices that communicate with legacy industrial communication standards and may not be compatible with current cryptographic services or blockchain smart contracts. Because of this reason, the proposed architecture takes the form of a gateway-based integration model. Within this model, legacy field devices communicate via their native protocols (e.g., Modbus/TCP, DNP3, OPC UA, or vendor-specific interfaces), and an intermediary edge or protocol-adaptation gateway receives telemetry and command data and forwards it to the anomaly-detection and blockchain layers. Device authentication, hash-based encryption, and interaction with smart contracts are handled by the gateway or edge server, not the field device itself. Edge IIoT layer The lower tier of the system consists of a group of IIoT-enabled sensors and actuators that help to measure the level, flow, pressure, and water quality in the distribution network. Depending on the installation, such devices are linked via industrial field buses or a low-power wireless protocol. Each device has a cryptographic identity that is stored in secure hardware, and measurement packets are to be originated by authenticated sources. An edge computing node transforms raw measurements into edge computing messages, cleanses them of noise, checks for data inconsistencies, attaches signatures to outgoing messages using lightweight cryptographic algorithms, and sends them to the communication middleware. This ensures that data integrity and authenticity remain intact during generation and that there is no risk of adversaries injecting fake values or infiltrating unsecured signals. Blockchain layer On top of the device layer, the architecture uses a permissioned blockchain network to ensure data integrity and generate a transparent, tamper-resistant audit trail. Fast block confirmation can be achieved at relatively low computational cost through a Byzantine-fault-tolerant consensus group of PBFT validator nodes distributed across gateways, SCADA servers, and cloud elements. Upon reaching a gateway, sensor data is hashed and sent as a ledger transaction, including metadata such as the time lag (timestamp), device identity, and anomaly score. The whole process ensures that all reading and operational events have an absolute record, with forensic reconstruction, and therefore, past logs cannot be changed by malicious actors. As shown in Algorithm 1, the proposed permissioned blockchain layer validates IIoT telemetry through device authorization, timestamp freshness checking, gateway signature verification, smart contract-based rule enforcement, and PBFT-based block confirmation before anchoring the telemetry digest to the ledger. Algorithm 1: Permissioned Blockchain-Based Integrity Anchoring and Smart Contract Validation for IIoT Telemetry. As shown in Fig. 2, the blockchain layer connects to IIoT devices. The lower tier of the figure illustrates how raw sensor measurements are transmitted from field IIoT devices to edge gateway nodes, where each reading is cryptographically hashed and encapsulated as a signed transaction. These transactions propagate upward to a distributed layer of blockchain validator nodes, which collectively execute a consensus algorithm. In either case, arrows can show both the addition of data and verification responses, visually highlighting how sensor measurements would be cryptographically anchored in the registry. Blockchain layer integration with IIoT Devices. Smart contract structure The blockchain administers smart contracts that implement integrity checks and automate security-related decisions. When a device submits a transaction to a gateway, the smart contract verifies the device’s identity and signature, ensures the timestamp is not too old, and confirms that it conforms to the scheduled reporting schedule. It also assesses the measurement’s compliance with operational constraints based on historical norms or hydraulic rules. If the transaction meets the specified conditions, the value is added to the ledger; otherwise, the contract triggers an alert that is sent to the anomaly-detection engine. Figure 3 illustrates this workflow as a path of sequential verification that starts with incoming sensor values, then proceeds to identity and signature verification, timestamp validation, operational rule verification, and the final operation: ledger write. The graph must indicate the conditional bifurcation at which violations trigger the creation of a security warning or anomaly flag, reflecting the deterministic, automated character of contract execution. Smart contract workflow for data integrity verification. Communication security protocols The sensors, gateways, SCADA servers, and analytics platforms communicate over secure transport protocols suitable for the network in use. TLS is used to encrypt IP-based traffic, while the lightweight DTLS is used to encrypt small IIoT links. The use of industrial WAN connections, based on MACsec or virtual private networks, to secure segments in the broad area. Fieldbus protocols are directly encased by secure variants where feasible, e.g., Modbus-TLS or OPC-UA authenticated sessions. X.509 certificates are used to verify each other through mutual authentication or blockchain-supported device identities. Replay attacks are avoided, and signed sequence counters provide time-based validation of packets, treating the communication exchange as reliable. Data flow across layers At the IIoT sensing layer, data is generated, sensed, and locally encrypted, then extracted and processed. These marked measurements flow through the communication layer to the SCADA systems, which use real-time data streams to regulate the processes at hand continuously. Simultaneously, a parallel data path streamlines the hashing of summaries for every reading into the blockchain network and establishes an irrevocable integrity anchor. This anomaly-detection engine is used based on the real-time feed and on ledger data that has been proven historically to cross-check physical conduct and cyber-consistency. The detection modules, SCADA, and operator dashboards should include feedback loops to ensure that warning or alarm indicators are visible and that advisories against actions taken in operations can be implemented as soon as needed. Integration with anomaly detection system The anomaly-detection system is deeply integrated into the architecture to ensure sustained situational awareness. It examines multivariate sensor streams, identifies abnormalities in normal hydraulic and cyber behavior, and produces real-time anomaly scores. These outputs are associated with blockchain-authenticated measurements, which provide additional confidence in the detection results. Anomalies detected on the edge node can be configured in real time, while those detected on the cloud or SCADA layers can benefit from additional computing resources and historical data. Alarms cannot be changed, as essential detections are recorded to the blockchain through smart contract systems. It offers a restrictive security environment where the creation, verification, analysis, and data auditing are highly synchronized. Blockchain-enabled anomaly detection framework The anomaly-detection framework, which is based on blockchain and incorporates both physical and cyber data streams across multiple layers, as well as verification mechanisms that cannot be altered, can reliably detect malicious or abnormal behavior in water-distribution CPS. The anomaly-detection model works across coordinated cyber and physical observations taken on the IIoT, communication, and control layers. The physical observation vector, at every time instant t, is denoted by: and includes process variables such as pressure, flow, tank level, water-quality indicators, and actuator states. After timestamp alignment, missing-value handling, and feature scaling, the normalized process vector is written as: where denotes the preprocessing operator. To capture temporal dynamics, a sliding window of length wis formed: with . Simultaneously, the features of the cyber layers are obtained from the existing communication- and control-related metadata for each observation window. Let represent the feature of cyber at the time t, and p is the size of the cyber-layer features. The characteristics are determined in the current study on the basis of control and event metadata present in the benchmark data representation, including the frequency of command occurrences, the patterns of control-state transitions, timing irregularity indicators, and other consistency-related descriptors in the same time window. The fused representation used by the detector is then written as: where denotes the vectorized form of the physical-process window , and denotes concatenation. The proposed model, therefore, combines a temporal encoder–decoder with cyber–physical feature fusion. Let denote the LSTM autoencoder parameterized by , and let be the reconstructed window. The reconstruction error at time t is defined as: where is the Frobenius norm. A cyber inconsistency score is computed from the auxiliary cyber vector using a lightweight scoring function . In the present implementation, is defined as the standardized distance of from its validation-set mean: where and are estimated from normal validation data. The final anomaly score is calculated as a weighted combination of the physical reconstruction error and the cyber inconsistency score: where and the minmax normalized versions of the physical reconstruction error and cyber inconsistency score respectively, such that the two contributions are on a similar scale. Here, is a relative weight of the physical and cyber layer, with indicating the utilization of the physical reconstruction term alone, corresponds to using only the cyber scoring term, and intermediate values represent fused decision-making. An anomaly label is then assigned according to: where is selected from the validation set, while the effect of is examined through ablation and sensitivity analysis in the revised experimental study. For integrity preservation, each accepted telemetry window is hashed before ledger submission: and are anchored to a permissioned blockchain. Other design features. The design is not based on storing raw process data on-chain, but rather on storing compact integrity data and anomaly evidence on-chain, thereby minimizing ledger overhead while maintaining auditability. This is why the implementation of the blockchain layer complements the learning-based detector, but is not its replacement. An anomaly model can be used off-chain on edge or near-edge compute devices, but the blockchain layer requires devices to authenticate, provide event provenance, and enable tamper-evident logging of security-relevant outputs. Experimental setup The suggested secure IIoT architecture, along with the blockchain-based anomaly detection scheme, has been tested using a combination of physical computing assets, virtualized environments, and containerized services to enable reproducibility and controlled experimentation. Its overall structure combines machine-learning pipelines, blockchain nodes, and real-time data-ingestion systems to mimic the operational patterns of water-distribution cyber-physical systems. Hardware and software configuration The experiments have been performed on a workstation equipped with a multi-core Intel Xeon processor, 64 GB of RAM, and an NVIDIA RTX-series graphics card to train deep learning models. The blockchain edge implementation was deployed on several lightweight virtual machines with 4 vCPUs and 8 GB of memory, which serve as distributed validator nodes for a permissioned ledger. The machine learning libraries TensorFlow and PyTorch (both based on Python) and Scikit-learn (classical machine learning baselines) were available in the software stack. The blockchain network was implemented in a Hyperledger Fabric-style environment with a low-latency setup. Docker containers provide a way to isolate pipelines, blockchain nodes, and data ingestion modules, allowing them to be tested simultaneously under identical conditions. Dataset preprocessing and data split strategy The data utilized in this study include the Battle of the Attack Detection Algorithm (BATADAL)47, Water Distribution (WADI)48, and SWaT49. Cleaning of the BATADAL, WADI, and SWaT data sets was done according to the same pipeline and standard processing to achieve uniformity and a fair comparison of models. First, the times were extracted and sorted in chronological order. Deduplicated data was performed, and all sensor/actuator channels were ordered in the same sequence. For variables that were continuous, such as process variables, linear interpolation was applied in cases of small gaps, and forward filling was applied for actuator and device status variables. Records were removed from the model training set if there were missing data for longer than 12 h or if the entries were not contiguous. All other variables, such as pressure, flow, tank level, conductivity, and other process variables, were normalized with min-max scaling based only on the training set. Same scalings as before were also used for the validation and test sets to prevent data leakage. Numerical state indicators represented the binary and categorical actuator states. For each temporal window, cyber layer descriptors were extracted, such as command transition frequency, number of actuator state changes, timing irregularity indicators, repeated value ratio, and consistency measures of communication. The temporal input sequence was built by using a sliding window of length τ. The τ values were chosen in the experiments from the validation performance of candidate values {10, 20, 30, 50}. The last one was chosen after optimizing the validation F1 score and consistency in the reconstruction loss. The LSTM autoencoder is trained during normal operating periods since it takes in the temporal behavior without the attack. The validation set was employed to early stop, threshold calibrate, and select the fusion weights. Only normal intervals and attack intervals were used for the test set, and they were only used for the final measurement of performance. In order to prevent the leakage of information, all splits were done sequentially, not randomly. All the training samples were normal; some calibration windows were included in the validation samples, and both normal and attack periods were included in the test samples. All models, SVM, Random Forest, LSTM Autoencoder, and the proposed hybrid model were subjected to the same preprocessing, window construction, and scaling protocol. Hyperparameter optimization The hyperparameters of the machine-learning and deep-learning models were trained on the validation set using grid-based search and cross-validation. The parameters to be adjusted were the learning rates, batch sizes, temporal window, and the sizes of the hidden units in the LSTM and the autoencoders. The optimal criterion used was to minimize reconstruction error or to maximize the distance between the normal and anomalous distributions. To control overfitting, dropout and early stopping were used for regularization, and loss-based metrics were used to assess model convergence. The same validation structure was used during training to enable a fair comparison of classical ML baselines. All baseline models were trained, validated, and tested using the same procedure to provide a fair comparison. The validation set was used for hyperparameter tuning, and only the test set was used for final metrics. To compare the proposed model with the classical machine learning approach, the window-level feature vectors were extracted from the same set of normalized input channels and fed to the classical machine learning models. For sequence-based models, all models used the same time window and time split. A validation loss was introduced to deep learning models to allow for early stopping, and all methods were evaluated using the same metrics. Blockchain deployment setup The proposed device is a hybrid LSTM autoencoder tasked with simulating the normal temporal behavior of water-distribution CPS signals. In physical-process variables, where n is the length of the sliding window, and w is the width of the sliding window (a normalized quantity), let be a sliding window. The encoder is made up of stacked LSTM layers, which convert the input sequence into a small latent representation h t, and the decoder recovers the original sequence using this latent state. The model is trained only on normal-operation data to ensure that the occurrence of abnormal conditions results in a higher reconstruction error. Simultaneously, an extracted feature vector of the network is obtained over the same period, that is, communication-rate, timing, retransmission, and command-pattern features. In the current design, both physical and cyber information are merged at the anomaly-scoring level, rather than being merged physically via direct concatenation within the recurrent encoder. This enables the temporal autoencoder to capture the dynamics of physical processes, and the cyber branch captures communication-side anomalies relevant to attacks and misuse of control. The model is also trained to minimize the reconstruction loss in normal windows. The output of the decoder can be denoted by . The reconstruction goal is formulated as: where B is the batch size, w is the temporal window length, and n is the number of physical variables. The current implementation, therefore, optimizes on the mean-squared reconstruction error. Optimization of the model on normal-operation training data is done through Adam-based gradient descent. Validation loss is used to perform early stopping to reduce overfitting. Anomaly detection (via reconstruction deviation) is performed in the proposed framework rather than supervised classification of attacks, so there is no additional temporal prediction loss in the present implementation. Evaluation metrics The performance evaluation plan used detection accuracy, recall, precision, and the F1-score (the harmonic mean of the two) to assess the models’ ability to distinguish between abnormal behavior and regular operations. Overall separability of thresholds was assessed using the Area Under the ROC Curve (AUC), and the false-positive rates were evaluated through their practical application to water-distribution systems. The measures that were monitored in the blockchain assessment included average transaction latency, block finalization time, and the cost of smart contract computation. The combination of these measurements provided detailed insight into the functionality of the integrated architecture as both a security device and a working part of a real-time CPS landscape. Results and discussion Performance on BATADAL dataset The BATADAL dataset provides a controlled setting for analyzing the dynamics of anomaly-detection models under hydraulic disturbances and sequences of cyber and physical attacks. The comparison of the models is shown in Fig. 4, with BATADAL detection performance in terms of ROC and Precision-Recall (PR) behavior. As Fig. 4(a) demonstrates that the Hybrid LSTM-AE, which offers cyber-physical feature fusion, yields the most robust ROC profile, and its curve is closer to the top-left area than the other baseline approaches, which is important because it indicates that the proposed algorithm offers a better trade-off between false-positive rate and true-positive rate. Figure 4(b) shows a steady trend in which the proposed model is more accurate across most of the recall range, even though class imbalance in anomaly-detection tasks is a problem. To enhance statistical robustness, the amended manuscript supplements these curve-based findings with variability-conscious reporting, including repeated executions of the results in the discussion and summary statistics. Multivariate detection performance on BATADAL. This is supported by the metrics that are provided in great detail in Table 4. The hybrid model achieves an accuracy of 0.97, a recall of 0.93, and an F1-score of 0.94, all significantly higher than those of the baselines. These results demonstrate that temporal dynamics, in combination with cyber-layer indicators, have meaningful positive effects on hydraulic networks for anomaly detection. Table 4Detection Metrics on BATADAL Dataset.ModelAccuracyRecallPrecisionF1-ScoreAUCSVM0.890.820.840.830.91Random Forest0.920.860.880.870.94LSTM Autoencoder0.940.900.920.910.96Hybrid LSTM–AE (Proposed)0.970.930.950.940.98 Detection Metrics on BATADAL Dataset. Performance on WADI dataset WADI is much more complex because it is based on a real-world testbed scenario with several pumps, tanks, and components, all controlled by a PLC. The reconstruction-error distributions in Fig. 5(a) indicate that the hybrid model can capture deviations effectively in the case of cyber-physical attacks. Attack windows of multi-order-magnitude reconstruction error generate a distinct separation between normal distributions. The classification behavior can also be explained by Fig. 5(b) using the confusion matrix. Specifically, the model rightly classifies 320 attack cases as true positive (TP) and 1155 normal cases as true negative (TN), and 45 false positives (FP) and 30 false negatives (FN). These values mean that the suggested approach has a rather low false-alarm rate and, at the same time, it identifies the majority of attack events, which is particularly crucial in CPS settings where unnecessary alarms and missed attacks can have crucial operational implications. Anomaly-detection behavior on WADI. Table 5, which summarizes the detection metrics, indicates that the proposed hybrid model has the highest recall of 0.92 and F1-score of 0.91 on WADI and is significantly better than the baselines. These findings demonstrate that it can identify small-scale, multi-stage intrusions into distributed water networks. Table 5Comparative Performance Results on the WADI Dataset.ModelAccuracyRecallPrecisionF1-ScoreAUCSVM0.840.760.800.780.87Random Forest0.880.810.850.830.90LSTM Autoencoder0.910.870.890.880.93Hybrid LSTM–AE (Proposed)0.940.920.900.910.95 Comparative Performance Results on the WADI Dataset. Performance on SWaT dataset The SWaT dataset is a continuation of the analysis of a multi-stage chemical treatment plant with complicated interactions at six operational stages. The scores on temporal anomalies in Fig. 6(a) indicate that the proposed model identifies attack intervals early, even in sequences intended to replicate natural changes. The sensitivity of the model to the cross-stage dependencies violations is also demonstrated in Fig. 6(b), wherein the variables of chemical dosing flow rate, conductivity, and pressure are strong contributors to the formation of the anomaly-score. Evaluation of SWaT under diverse attack types. The overview in Table 6, indicates that the hybrid approach attains high sensitivity (0.94) and specificity (0.95), which are far superior to those of classical ML algorithms. This supports the generalization of the framework to other industrial CPS environments beyond water distribution, such as energy grids and manufacturing systems. Table 6Detection metrics on the SWaT dataset.ModelAccuracyRecallPrecisionF1-ScoreSVM0.840.790.830.81Random Forest0.880.840.880.86LSTM Autoencoder0.920.900.920.91Hybrid LSTM–AE (Proposed)0.950.940.940.94 Detection metrics on the SWaT dataset. To measure the operational overhead of the permissioned blockchain layer, we measured transaction-confirmation latency, block-finalization latency, and sustained throughput under varying workload conditions. At moderate load, the blockchain was found to have a sub-second average confirmation latency, and higher transaction rates caused both average and tail latency to increase due to the coordination and overhead of consensus messages among validators. Ablation and sensitivity analysis of cyber physical fusion To validate the weighted fusion approach conducted in Eq. 15, some ablation and sensitivity analysis is performed by changing the physical reconstruction weight α. The complementary cyber inconsistency weight was defined as . Hence, stands for physical LSTM autoencoder, stands for cyber inconsistency branch, and the values in between indicate cyber-physical fusion. The validation set was used to determine the operating threshold for each setting, and the test set was left unobserved during the process of threshold calibration. The cyber-only score can help identify communication and command irregularities, as demonstrated in Table 7, but is not as sensitive to slow hydraulic deviations. The physical-only model can describe changes in time in a physical process, but can’t describe attacks that are first detected as inconsistencies in cyber layers. The best overall performance is achieved when both sources of evidence are fused, thus demonstrating the utility of complementary physical and cyber evidence in the proposed anomaly score. Table 7Sensitivity Analysis of the Fusion Weight in the Proposed Cyber Physical Anomaly Score. Fusion settingAccuracyRecallPrecisionF1 scoreAUC0.001.00Cyber score only0.890.840.860.850.900.250.75Cyber-dominant fusion0.920.890.880.880.930.500.50Balanced fusion0.950.920.930.920.960.750.25Physical dominant fusion0.960.930.940.930.971.000.00Physical reconstruction only0.940.900.920.910.96 Sensitivity Analysis of the Fusion Weight in the Proposed Cyber Physical Anomaly Score. From the results in Table 7, it can be seen that the fused settings work better than either of the single-branch settings. Physically dominant fusion appears to give the most balanced fusion since the temporal hydraulic reconstruction error reflects the deviations from the process, and the cyber inconsistency score provides early information from command frequency, command timing irregularity, and changes in device state. According to this trend of validation, the weights of and were used in the final experiments. This decision was not made by hand – it was done through a sensitivity analysis that was validated. Interpretability analysis for operator trust However, water distribution CPS is a safety-critical environment, so the accuracy of detection is not enough for operational trust. The new framework, thus, introduces an interpretability layer that splits each anomaly score into a physical reconstruction contribution and a cyber inconsistency contribution. In the physical branch, reconstruction errors are calculated level-wise, sensor channel-wise for every feature detected in the detected window. The dominant physical contributors to the anomaly are considered to be the variables with large reconstruction deviations. In the cyber branch, the standardized distance of each cyber descriptor from the normal validation distribution is employed to determine if the anomaly is correlated to the frequency of commands, deviation from the timing of commands, state transition of devices, or communication consistency deviation. This interpretation process assists the operators in identifying process faults from communication-driven attacks. A significant reconstruction error for tank level, pressure, and flow indicates that the hydraulic process is behaving outside of expected normal behavior. A high cyber inconsistency contribution and moderate physical reconstruction error, on the other hand, indicates an early-stage cyber manipulation, replay, or command sequence irregularity before a large hydraulic impact. The anomaly label, confidence score, dominant physical variables, dominant cyber descriptors, and the blockchain transaction identifier associated with the verified telemetry window are then displayed to the operators in the final alarm. This enhances transparency of the decision by allowing the operator to see not only whether the event was anomalous, but which aspects of the process and communication led him to make this conclusion. The summary of the proposed explanation procedure is provided in Table 8, where no retraining of LSTM autoencoder is needed. It does so, however, by reusing errors of reconstruction and deviations in the cyber descriptors to provide operator-readable evidence. This helps to better meet the needs of critical water CPS environments where alarm justification, forensic review, and response prioritization are needed. Table 8Example operator-oriented interpretation of anomaly decisions.Detected event typeDominant physical contributorsDominant cyber contributorsInterpretation for operatorSensor manipulationPressure and flow reconstruction errorModerate timing deviationPossible false data injection affecting hydraulic readingsReplay attackLow to moderate process deviationHigh timestamp and sequence inconsistencyPreviously valid packets may have been resentActuator command abuseTank level and pump status deviationHigh command transition frequencyPossible unauthorized or abnormal control actionProcess faultHigh tank level or flow reconstruction errorLow cyber inconsistencyLikely physical fault rather than cyber intrusionCoordinated attackHigh physical reconstruction errorHigh command and timing inconsistencyCyber physical attack with observable process impact Example operator-oriented interpretation of anomaly decisions. Cross-dataset generalization analysis Cross-dataset experiments indicate that anomaly detectors can generalize beyond the environment in which they were trained. As seen in Fig. 7(a) and Fig. 7(b), the transferability differences are significant. Models trained on WADI, which are dense in sensors and physically diverse, should generalize to SWaT and perform moderately well in recall even without retraining. Conversely, models trained only on BATADAL generalize less well because they use a simpler hydraulic model and lack the complexity of a cyber-layer. Cross-dataset robustness evaluation. This asymmetry is pointed out by the quantitative measures in Table 9. WADI training achieves the best cross-dataset performance, while BATADAL training achieves the worst, indicating the need for a more diverse training set to better reflect real-world conditions. Table 9Cross dataset generalization results under transfer evaluation.Training DatasetTesting DatasetRecallF1-ScoreAUCBATADALWADI0.610.580.72BATADALSWaT0.550.540.70WADIBATADAL0.780.750.85WADISWaT0.820.800.88SWaTWADI0.760.730.84 Cross dataset generalization results under transfer evaluation. Blockchain performance and overhead assessment The performance of blockchains was assessed to determine whether integrity anchoring and smart contract validation introduce operational delays of a material level that compromise real-time anomaly detection. The trends in block-finalization latencies in Fig. 8(a) indicate that the permissioned ledger can be used for operational logging and auditing under moderate load, as it can confirm within less than a second. Latency also increases linearly as load increases, because consensus messages incur overhead in non-blocking integrity verification. The throughput behavior presented in Fig. 8(b) indicates that the ledger scales well with additional validator nodes until reaching network saturation. Blockchain performance characteristics. Comparison with state-of-the-art methods A rigorous comparison was conducted between the proposed hybrid blockchain-enabled anomaly detection framework and state-of-the-art approaches reported in recent literature on water CPS security, ICS intrusion detection, and deep-learning–based anomaly detection. The selected works represent diverse methodological categories, including statistical rule-based detectors, subspace-learning approaches, deep temporal architectures, graph-theoretic CPS models, and hybrid cyber–physical intrusion detection systems. Table 10 shows the average latency and throughput performances of the implemented permissioned blockchain with different workloads. As transaction load rises, confirmation and finalization delays grow as a result of increasing communication and coordination burden between the validators, and as the load on the system falls short of its saturation point, the throughput of transactions continues to rise. The implication of these results is that they are intended to show the overhead of the implemented blockchain integrity layer, and are not intended to be compared against a no-blockchain baseline or other consensus mechanisms. Table 10Model sensitivity and specificity on SWaT.ConditionMean latency (ms)P50 (ms)P95 (ms)P99 (ms)Finalization latency (ms)Throughput (tx/s)Moderate load≈ 350–450≈ 320–400≈ 600–750≈ 800–950≈ 500–700≈ 40–60Higher load≈ 550–750≈ 500–650≈ 900–1200≈ 1200–1500≈ 800–1100≈ 60–80Near saturation≈ 900–1300≈ 850–1100≈ 1500–2000≈ 2000–2500≈ 1300–1800≈ 70–90 Model sensitivity and specificity on SWaT. Table 11 highlight clear performance differences across approaches. Classical threshold-based detectors, such as CUSUM and Shewhart-based methods, perform inconsistently under nonstationary hydraulic behavior, resulting in reduced recall when faced with subtle and long-duration attacks. Subspace-learning methods, such as PCA-based detectors, improve the structural understanding of sensor correlations but are vulnerable to concept drift and cyber-induced divergence from linear subspaces. Deep-learning approaches, such as LSTM-based IDS and CNN–LSTM hybrids, exhibit significantly improved temporal sensitivity and higher AUC values, particularly in multistage testbeds such as WADI and SWaT. However, these methods typically lack mechanisms to guarantee data integrity or to support auditability after attacks. CPS anomaly detectors are graph-based (e.g., GNN-based smart water models, which find stage-specific detection well because the hydraulic topology and inter-sensor relationships are modeled, but at a high cost). Table 11Benchmark comparison with recent literature.ReferenceBlockchain SupportPrimary Dataset(s)Evaluation BasisAccuracyRecallF1-ScoreAUCComparability Note 50 NoSWaTReported from original publication0.930.880.900.95May use different preprocessing and split settings 51 NoWADIReported from original publication0.940.900.910.96Direct comparability may be limited 52 NoBATADALReported from original publication0.900.850.870.92Dataset protocol may differ from this study 53 NoSWaTReported from original publication0.910.860.880.93Metrics based on prior reported setup 54 NoWADIReported from original publication0.920.880.890.94Not a strictly controlled head-to-head comparisonProposed Hybrid LSTM–AE + Cyber–Physical Fusion + BlockchainYesBATADAL, WADI, SWaTEvaluated in this study under the present preprocessing and split policy0.950.920.930.97Includes blockchain-based provenance in addition to detection Benchmark comparison with recent literature. Conversely, the Hybrid LSTM -AE based on cyber-physical features fusion and blockchain-supported provenance is always superior to the literature in accuracy, recall, F1, and AUC, as evident in Table 11. The tamper-resistant data logging method offered by the integrated blockchain layer enables cross-validation of anomalies and forensic reconstruction, features not implemented in other methods. Such a set of advantages has enabled the proposed solution not only to achieve state-of-the-art detection quality but also to be implemented safely and transparently in contemporary water distribution CPS. Conclusion and future work This paper presents an integrated design for a secure IIoT system for water-distribution CPS, including a blockchain-based data-integrity mechanism and a hybrid anomaly-detection model that detects both cyber and physical deviations. The suggested system includes secure IIoT measurements, tamper-resistant blockchain recording, smart-contract enforcement, and multi-layer anomaly detection based on temporal deep learning and cyber-physical feature fusion. The power and versatility of the framework were demonstrated through extensive experiments on the BATADAL, WADI, and SWaT datasets across various hydraulic and industrial systems. Verifiable provenance and transparent auditing, which were also a bane of critical water infrastructure, especially regarding trust and data authenticity, were introduced by the blockchain layer. Experiments showed that the proposed hybrid LSTM-AE architecture was consistently more effective than classical machine learning approaches and other available deep learning baselines. Across all datasets, it performed better on recall and F1 scores and was more sensitive to nuanced disturbances and multi-step attack schemes. The cross-dataset tests indicated that models trained in more complex and sensor-diverse environments (such as WADI) performed better in transfer than models trained in simpler hydraulic simulations. This validates the strong link between source domain diversity, feature compatibility, and attack pattern similarity with respect to cross-dataset robustness. It was discovered that the blockchain-based validation pipeline was functional in pegging telemetry to an unalterable registry at a controllable latency cost, so that anomalies, along with data streams, could still be provenanced even in an adversarial situation. Taken together with the results above, they validate that deep-learning anomaly detection, combined with distributed ledger technologies, is a practical starting point for making modern water CPS secure. The limitations in the proposed system are few and far between, despite its high performance. Like most hybrid deep learning detection models, the model requires a large amount of historical data to train effectively, which is often unavailable in small or newly deployed water-distribution systems. Although blockchain offers tamper resistance, it introduces additional computational and communication overhead and can be constrained to use in edge cases with significant resource constraints. Another aspect of the current architecture is its reliance on monitored and semi-monitored training methods, which can make the model vulnerable to unknown or emerging attack types unless appropriate retraining mechanisms are in place. One drawback is that domain shift is still a problem for cross-dataset generalization. The sensing density, the process topology, the actuator logic, the sampling behavior, and the attack design are different aspects that differentiate BATADAL, WADI, and SWaT. Thus, a threshold and representation discovered from one data set may not be a perfect match to another operational domain. This cyber-physical fusion is useful for making the system robust; however, to deploy the system in a new utility network, local threshold recalibration, target domain validation, or domain adaptation should be implemented. Future research: It is possible to investigate adaptive, self-evolving anomaly-detection models that continuously learn from streaming data, thereby reducing the need for large-scale offline retraining. The other direction is to combine physics-informed neural networks and digital-twin simulation frameworks to create more interpretable schemes and offer scenario-driven analysis of attack outcomes. It will also seek to expand the blockchain to lightweight consensus protocols, hierarchical ledgers, or edge-friendly verification schemes to achieve low overhead and enhance scalability. Long-term initiatives can also involve developing federated or privacy-preserving learning systems that allow the sharing of cross-utility knowledge without necessarily compromising the secrecy of their operations. Finally, extending the framework to other domains, such as wastewater systems, district heating, and multi-utility CPS, will help create a generalized security framework for the next generation of connected, interdependent critical infrastructures. Acknowledgements This study was partially carried out within the project MOST – Sustainable Mobility National Research Center and received funding from the European Union Next-GenerationEU (PIANO NAZIONALE DI RIPRESA E RESILIENZA (PNRR) – MISSIONE 4 COMPONENTE 2, INVESTIMENTO 1.4 – D.D. 1033 17/06/2022, CN00000023). Author contributions Abdu Salam: Conceptualization; Methodology; System architecture design; Writing – original draft.Ahmed Thair Shakir: Formal analysis; Investigation; Validation; Data interpretation. Qaisar Javaid: Formal analysis; Investigation; Writing – review and editing. Barbara M. Masini: Methodology; Validation; Writing – review and editing. Masood Ahmad: Software; Data curation; Visualization. Ishtiaq Wahid: Supervision; Resources; Project administration; Writing – review and editing. All authors have read and approved the final manuscript. Funding This study was partially carried out within the project MOST – Sustainable Mobility National Research Center and received funding from the European Union Next-GenerationEU (PIANO NAZIONALE DI RIPRESA E RESILIENZA (PNRR) – MISSIONE 4 COMPONENTE 2, INVESTIMENTO 1.4 – D.D. 1033 17/06/2022, CN00000023). Data availability The datasets analyzed during the current study are publicly available at[https://www.batadal.net/data.html](https:/www.batadal.net/data.html), [https://service.tib.eu/ldmservice/dataset/swat-and-wadi-datasets](https:/service.tib.eu/ldmservice/dataset/swat-and-wadi-datasets) and[https://itrust.sutd.edu.sg/itrust-labs\_datasets/dataset\_info/](https:/itrust.sutd.sutd.edu.sg/itrust-labs_datasets/dataset_info). Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Ahmed Thair Shakir, Email: ahmedthair0@uoanbar.edu.iq. Barbara M. Masini, Email: barbaramavi.masini@cnr.it References Associated Data Data Availability Statement The datasets analyzed during the current study are publicly available at[https://www.batadal.net/data.html](https:/www.batadal.net/data.html), [https://service.tib.eu/ldmservice/dataset/swat-and-wadi-datasets](https:/service.tib.eu/ldmservice/dataset/swat-and-wadi-datasets) and[https://itrust.sutd.edu.sg/itrust-labs\_datasets/dataset\_info/](https:/itrust.sutd.sutd.edu.sg/itrust-labs_datasets/dataset_info).