Machine Learning on Blockchain (MLOB): A New Paradigm for Computational Security in Engineering

Zhiming Dong; Weisheng Lu

doi:10.1016/j.eng.2024.11.026

Engineering ›› 2025, Vol. 47 ›› Issue (4) :267 -281. DOI: 10.1016/j.eng.2024.11.026

Research Engineering Management—Article

research-article

Machine Learning on Blockchain (MLOB): A New Paradigm for Computational Security in Engineering

Zhiming Dong
, Weisheng Lu ^*

Author information +

History +

PDF (3156KB)

Abstract

Machine learning (ML) has been increasingly adopted to solve engineering problems with performance gauged by accuracy, efficiency, and security. Notably, blockchain technology (BT) has been added to ML when security is a particular concern. Nevertheless, there is a research gap that prevailing solutions focus primarily on data security using blockchain but ignore computational security, making the traditional ML process vulnerable to off-chain risks. Therefore, the research objective is to develop a novel ML on blockchain (MLOB) framework to ensure both the data and computational process security. The central tenet is to place them both on the blockchain, execute them as blockchain smart contracts, and protect the execution records on-chain. The framework is established by developing a prototype and further calibrated using a case study of industrial inspection. It is shown that the MLOB framework, compared with existing ML and BT isolated solutions, is superior in terms of security (successfully defending against corruption on six designed attack scenario), maintaining accuracy (0.01% difference with baseline), albeit with a slightly compromised efficiency (0.231 second latency increased). The key finding is MLOB can significantly enhances the computational security of engineering computing without increasing computing power demands. This finding can alleviate concerns regarding the computational resource requirements of ML–BT integration. With proper adaption, the MLOB framework can inform various novel solutions to achieve computational security in broader engineering challenges.

Graphical abstract

Keywords

Engineering computing / Machine learning / Blockchain / Blockchain smart contract / Deployable framework

Cite this article

Download citation ▾

Zhiming Dong, Weisheng Lu. Machine Learning on Blockchain (MLOB): A New Paradigm for Computational Security in Engineering. Engineering, 2025, 47(4): 267-281 DOI:10.1016/j.eng.2024.11.026

登录浏览全文

4963

注册一个新账户忘记密码

1. Introduction

Computing methods, algorithms, and software play a vital role in engineering domains, including system and process design [1], simulation [2], control [3], and optimization [4]. Machine learning (ML) engages approaches that enable extraction of meaningful patterns from massive datasets and eliminate the need for explicit programming [5]. Compared to conventional computational methods, ML offers enhanced accuracy and efficiency [6]. It exhibits remarkable scalability in terms of accuracy, capable of tackling intricate engineering problems by learning highly nonlinear representations from extensive computational and experimental data amassed [7]. ML processes can also operate autonomously and adapt to new data and changing conditions [8], without the requirement for human interventions, thereby showing computational efficiency.

Nevertheless, ML-based computing faces security threats, encompassing data tampering, logic corruption, or both [9], and leading to potential manipulation, sabotage, compromised performance, and safety concerns [10]. The opaque nature of ML exacerbates these concerns, raising issues of security and accountability, which are crucial in specific engineering computing tasks like digital forensics [11], process control [12], quality assurance [13], and abnormality detection [14].

Blockchain, a decentralized ledger technology with the advantages of decentralization, transparency, immutability, and traceability [15], has been vigorously explored in the face of such security threats. Blockchain technology (BT) has been widely adopted by engineers and researchers to safeguard engineering data security, such as critical engineering data, intellectual property, and sensitive information related to engineering systems [16]. In such paradigm, BT is used to store transactions (e.g., input data, records, and computing results) in a decentralized, chained fashion to ensure data security and prevent tampering.

Nonetheless, such paradigm faces computational security risks caused by logical corruption. The ML models for engineering computing are still executed in an off-chain environment and handled by external agents, presenting clear risks for hackers to manipulate model parameters [17] or inject malicious code [18] to produce adversary-selected outputs [19]. In view of these risks, blockchain, particularly its smart contract (BSC), provides enormous potentials to protect both data and computational security. However, such potentials have not been well harnessed in the literature.

This research endeavors to construct a deployable framework combining ML and BT to enhance the computational security in engineering. The fundamental principle of this ML on blockchain (MLOB) framework is to embed all engineering computing records and transactions, including data, computing processes, and results, on the blockchain with minimal or no human intervention. To illustrate the MLOB framework, a prototype has been developed based on a real-life case. While such framework enhances computing security, it may incur costs in terms of efficiency and accuracy. This phenomenon can be called as the triangle of accuracy, efficiency, and security. It means that one needs to balance the three indicators in gauging the performance of a computing process.

The main contributions of this paper are summarized as follows.

•An ML–BT integrated framework named MLOB to ensure both data and computational security. The framework will guarantee that the data input, ML computing process, and result output are all conducted on the blockchain, and eliminates potential risks associated with exposure to the off-chain world.

•An accuracy, efficiency, and security “balance triangle” to evaluate the performance of MLOB. The triangle will provide a scientific explanation for the trade-off among accuracy, efficiency, and security for successfully implementation and future research objectives.

The remainder of the paper is organized as follows. Subsequent to this introductory section is a literature review of existing works on ML for engineering computing, the integration of BT and ML, and the accuracy, efficiency, security “triangle.” Section 3 illustrates the MLOB framework. The case study and experiment results are presented in Section 4 and Section 5, respectively. Section 6 discusses the potential of alternative paradigms, and explores future direction, while conclusions are drawn in Section 7.

2. Literature review

2.1. Machine learning for engineering computing

ML has been applied to computing tasks to solve engineering problems at both operation and system level [20]. Operation-level ML applications aim to optimize or reengineer individual engineering operations within a system. This involves replacing traditional computational methods with ML techniques to improve performance. For example, Guo et al. [21] proposed an intelligent clustering routing approach for unmanned aerial vehicle ad hoc networks to improve routing clustering efficiency, stability, and service quality. Yuan et al. [22] developed an accurate, building-level ML model to quantify urban material stock using machine learning and building features in Hong Kong, and analyzed the spatiotemporal evolution of building material stock [23]. Guo et al. [24] proposed a context-awareness trust management model for vehicular networks to evaluate the trustworthiness of received messages. Likewise, Dong et al. [25] proposed a boundary-aware transformer model for fine-grained composition recognition of construction waste mixtures, and Dong et al. [26] suggested a patch-based weakly supervised semantic segmentation network for crack detection in construction materials. Additionally, ML can be also used for biometric recognition tasks such as gait recognition [27] and face detection [28].

System-level applications of ML aim to optimize a collection of chained or networked operations within a system or inter systems. The purpose is to improve overall system performance, as defined by availability, productivity, energy efficiency, and quality. Teoh et al. [29] discussed predictive maintenance using genetic algorithm-based resource management and ML in fog computing. Lăzăroiu et al. [30] conducted a systematic review focusing on the Internet of Things (IoT) in manufacturing and its integration with deep learning-assisted smart process planning, robotic wireless sensor networks, and geospatial big data management algorithms. Li et al. [31] proposed a method to assess Window View Indices (WVIs) using a deep transfer learning method, which has potential applications in urban planning and development [32], [33].

These ML computing paradigms can be categorized as different scenarios as presented in Fig. 1. In Fig. 1(a), the ML models process the data input and then output the results, that is, in a centralized database or dashboard. This framework is reportedly effective. However, when the ML model is used in some specific engineering scenarios, human participants have an incentive to tamper with the input and output in the database for undesirable purposes [34]. Blockchain is introduced as a security technology under this circumstance.

2.2. Integration of machine learning with blockchain

The integration of ML and BT can be BT-centered or ML-centered [35]. The former entails using ML to analyze the data stored on BT to derive meaningful insights. Shinde et al. [36], for example, proposed an enhanced road construction management process, wherein ML algorithms are used for prediction and BT is used to ensure the safety of contracts among stakeholders. Wong et al. [37] explored ML for the scheduling and management of scarce resources, and BT is used to secure privacy data from cyber threats. Xu et al. [38], meanwhile, developed a blockchain-based framework named P-OSH to emphasize privacy protection in construction occupational safety and health (OSH) management. Kallapu et al. [39] proposed a blockchain-based attribute-aware encryption method for secure communication, and presents a new method for securely transmitting images over insecure channels in the latest work [40].

ML-centered integration involves use of BT to support information exchange during decentralized ML computing to eliminate information failures. Popular ML-centered integration focuses on training stage. Jovanovic et al. [41] explored the integration of blockchain and eXplainable Artificial Intelligence (XAI) in the context of federated learning for the improvement of financial credit scoring. Li et al. [42] proposed smart work package learning, which leverages a blockchain network to validate and store updated model parameters from each smart work package. Kasyap et al. [43] proposed a Byzantine-robust and inference-resistant federated learning framework that utilizes a permissioned blockchain and a tailored secure aggregation scheme to improve the overall performance.

Various data sources and datasets serve as the foundation for research on the integration of ML and BT. A significant portion of this research employed this integration to manage real-world engineering computing data [38], [42]. Meanwhile, for those studies striving to establish a theoretical framework, public datasets prove to be useful tools for performance evaluation [43]. Further, certain existing studies utilize qualitative methods such as surveys, questionnaires, and interviews for evaluation purposes [44].

Workflow of above works can be summarized as Fig. 1(b). Data input, and sometimes result output are stored on blockchain. However, after ML training, the ML computing for executing engineering computing task takes place externally, which is similar to the ML solutions in Fig. 1(a). While these computing paradigms have gained recognition for their effectiveness in safeguarding engineering computing data and records, potential computational security threats during the ML inference phase remain.

2.3. The accuracy, efficiency, security “triangle”

There is a fundamental tension between security, accuracy, and efficiency in engineering computing within limited computing resource, whereby improvement in one performance indicator is often at the cost of others. In real-life engineering computing tasks, a trade-off amongst the three performance indicators must be made. For example, Peethambaran et al. [45] proposed a privacy-based composite classifier model to balance accuracy and security, with its performance evaluated using a parallel computing framework. Cooper et al. [46] explored the trade-off between accuracy and efficiency in distributed ML systems, and discussed the interplays between distributed computing systems and ML algorithms. Al-Marridi et al. [47] proposed an intelligent and secure health system called Healthchain-RL, which can optimize the blockchain network’s behavior in real time to trade off security and efficiency.

For an ML–BT integrated framework, the incorporation of BT enhances security at the expense of accuracy and efficiency. Jiang et al. [48] proposed a blockchain-based model evaluation framework to address issues of untrustworthiness in traditional model evaluation approaches. Various verification mechanisms are introduced to reduce potential security threats from centralized model evaluation. Feng et al. [49] proposed CoBC, a blockchain-based collaborative inference system, to improve the sensing capabilities of individual IoT devices. The simulation results demonstrate CoBC’s good performance and practicality. Wang et al. [50] proposed SMARTINV, which is a framework that automates the detection of “machine un-auditable” bugs in smart contracts by inferring invariants across multiple modalities. Moreover, Adel et al. [51] proposed a decentralized ML system that utilizes BT as a computing-oriented technology for decision-making. These methods emphasized the security of ML-based engineering computing, but suffer from low execution efficiency, poor accuracy performance, or low scalability.

In summary, a research gap exists wherein current ML–BT integration solutions primarily focus on data security, but overlook computational security. Consequently, the research problem addressed in this study is ensuring both data and computational security during engineering computing. As depicted in Fig. 1(c), the proposed MLOB paradigm is developed for engineering computing and safeguards data input, record output, and the ML computing process in BT to ensure both data and computing security, meanwhile, achieving a trade-off amongst security, accuracy, and efficiency.

3. The proposed framework

The overall architecture of the MLOB framework, depicted in Fig. 2, allows ML-based computing logic to be executed on the blockchain, while also storing data input and output on the blockchain to ensure both data and computational security. The MLOB framework consists of four core components arranged in a linear sequence to achieve this objective. As illustrated in Fig. 2, the ML acquisition component trains a task-specific ML model, followed by ML conversion, which adapts the trained ML model for actual deployment. The third component loads the converted ML model onto the BSC in the blockchain. Finally, the ML-based computing task is executed on-chain. In more detail, Section 3.1 will cover the ML acquisition component, Section 3.2 will address ML conversion, Section 3.3 will discuss ML safe loading, and Section 3.4 will elaborate on consensus-based ML model execution.

3.1. Machine learning acquisition

ML acquisition is the process of obtaining an ML-based engineering computing process to perform a specific computational task. The first step, engineering computing logic design, aims to transform the traditional manual engineering computing logic into an automatic, ML-based executable logic. During this step, users must determine which parts of the original workflow can be implemented with ML [52]. Additionally, it is crucial to select the most appropriate ML models for the given problem.

The second step, model selection, refers to the process of selecting ML models based on specific engineering tasks. According to Bertolini et al. [53], the deployment of ML for engineering computing is challenging due to the overwhelming number of ML models provided in technical literature. A researcher must therefore carefully consider the characteristics of specific tasks in the ML model selection stage to choose a suitable model.

The final, crucial step is ML training. It aims is to train the chosen ML model by providing input data and comparing the predicted outputs with the actual outputs to minimize errors. Once the training is complete, assessing performance of the trained model will ensure that it can execute engineering computing logic with the expected accuracy and efficiency. However, it is important to note that the ML training procedure is not always mandatory. In some engineering computing tasks, existing open-access pre-trained models or general artificial intelligence models, such as ChatGPT and Segment Anything model (SAM) [54], can be utilized without training.

3.2. Machine learning conversion

Once obtained, the computing model undergoes conversion for safe execution on the BSC. This involves logic transfer and ML transversion. The former is the process of translating and rewriting the executable logic obtained above into a format that can be handled and identified by the BSC in the BT network.

ML transversion plays a crucial role in the ML conversion process. The architecture of ML models is often complex, and the trained ML weights are typically accompanied by well-designed data structures, hindering direct computing within the BSC. For example, some platforms store ML models as separate structures that contain model parameters, while others only export the model parameter files or maintain the architecture and weights separately. In the MLOB framework, TensorFlow.js [55] is utilized as an ML model execution engine in the BSC. Consequently, the ML model must be converted to a JavaScript-friendly format to be used in a BSC code execution environment.

3.3. Machine learning safe loading

The execution of ML-based engineering computing necessitates the loading of data and ML models into the BSC. To address the risks that may arise during the process of the data transmission between off-chain and on-chain world, ML safe loading is implemented. Algorithm 1 as shown in Table 1 illustrates the steps involved in the ML safe loading procedure, which include ML model loading, hash verification, and data loading.

To begin, the ML model and its corresponding hash value are loaded from a source outside the BT network, using a unique identifier such as a uniform resource locator (URL). As this procedure carries risks, hash verification is implemented to ensure the security of the ML model. When utilizing the model, the URL and hash are retrieved from the ledger in BT, and the model is downloaded via the provided URL. Subsequently, the hash of the downloaded model is retrieved and compared for verification, the verification result can be sent to both model user and uploader as a response. If a hacker were to modify the model file during the loading procedure, the hash of the modified file would differ from the stored hash and the user would be alerted to the fact that the model has been tampered with. For the process that uploading the data input to MLOB framework for on-chain processing, similar verification process is conducted to enhance the data security.

Once the loaded ML model successfully passes validation, the ML safe loading process proceeds to load the necessary data from the distributed ledger within the BT network. Finally, both the ML model and the data are loaded into the BSC runtime environment, completing all preparatory tasks before execution.

3.4. Consensus-based computing

Consensus-based computing execution involves executing engineering computing logic on blockchain and achieving consensus on its validity within a distributed framework. This mechanism plays a crucial role in ensuring the safety, correctness, and fault tolerance in a decentralized manner, creating a trusted environment that is resilient to attacks and failures, while also maintaining the security of the distributed ledger.

The consensus-based computing process is depicted in Fig. 3. Within the BT network, three types of nodes participate in this process: the client, peer, and orderer. The client serves as the gateway for organizations involved in the network to interact with it. Peers, on the other hand, play a fundamental role within the network as they belong to different organizations. They are responsible for storing the ledger and executing BSCs. Orderers are utilized for transaction ordering. Verified transactions undergo packaging into blocks through this ordering service, which are then stored in the distributed ledger.

The consensus-based computing process comprises:

(1) Initiation. The process initiates with the submission of the data as record, meticulously structured as a set of data including information pertaining to the engineering computing task, the record uploader’s identity, computing values, and associated attachments.

(2) Endorsing. A serious software development toolkit (SDK) acts as an intermediate processing unit, receiving this submission request and packaging it as a transaction proposal, complete with the contractor’s signature.

(3) Execution. The signed proposal is then propagated to all peers in the underlying BT network for execution and endorsement to get the proposal response with execution result and other necessary information. These endorsed transaction proposals are subsequently broadcast to the BT network.

(4) Verification. The clients of the organization then verify the consistency of the execution result and signatures against the endorsement policy. Once an adequate number of endorsement responses have been received, the proposal, along with the corresponding distributed ledger update request, is submitted for ordering. The orderer verifies the transaction information and signatures, after which the transaction is packaged into blocks. The final verification, carried out by the ordering service, ensures that the computing execution result is valid and ready to be written to the distributed ledger.

(5) Completion. After verification, the execution result of ML computing is written into the distributed ledger in each peer.

Any tampering attempt by a hacker, be it the modification of the BSC or manipulation of response result of a single peer, will lead to inconsistencies in the invoked results across other organizations. These inconsistencies can be identified by the BT network of MLOB. As a consequence, the transaction will be recognized as illegal.

4. An illustrative case

A prototype is developed and implemented in an actual engineering computing case: an indoor construction progress monitoring task. The objective of this task is to monitor and estimate the progress of indoor construction by comparing the as-is with as-built statuses. The as-is data consists of sequential images with corresponding poses (Fig. 4(a)), while the as-planned data is represented by the building information model (BIM) (Fig. 4(b)). Fig. 4 provides a visual representation of these components.

The construction progress monitoring process and its results are highly relevant to progress payments and quality accountability assurance [56]. However, the ML computing process is susceptible to potential threats or interference that can compromise the accuracy of the progress estimation process. This interference can arise from misunderstandings, conflicting interests, or miscommunication between the progress monitoring team and other stakeholders, and can lead to a failure in process control.

The primary objective of MLOB on this case is to incorporate the ML-driven progress monitoring workflow within the MLOB framework. All data input and result output must be stored on the blockchain network in a decentralized and distributed manner, while the ML-based progress monitoring workflow should be executed as a BSC to ensure transparent and secure ML computing execution. The deployment process should automate all facets, thereby eliminating the need for human intervention. The deployed MLOB in this case should possess the capacity to guarantee the security of both data and computations, encompassing measures to prevent tampering and corruption of data and computing logic, among other security considerations.

To achieve this objective, a prototype is implemented. Firstly, in the ML acquisition step, traditional indoor construction progress monitoring logic is implemented as a BSC, and a semantic segmentation ML model is trained using a deep learning tool called PyTorch. Secondly, in the ML conversion step, the trained ML model in .pth format is converted to a BSC-friendly format using the TensorFlow.js converter. Thirdly, in the ML safe loading step, the converted ML model is securely loaded onto the BSC through a meticulously designed security-enhanced loading procedure. Lastly, in the consensus-based computing step, the BSC for indoor construction progress monitoring is executed based on the ML model and input data, with the results stored as immutable records. Detailed explanations can be found in 4.1 Prototype implementation in the case, 4.2 Machine learning acquisition, 4.3 Machine learning conversion, 4.4 Machine learning safe loading, 4.5 Consensus-based computing.

4.1. Prototype implementation in the case

In this case scenario, the prototype is comprised of a computer with an Intel Xeon E5-2640 v4 CPU and 64 Gibibytes (GiB) memory, which offers ample processing power and storage capacity. The system runs on Ubuntu 18.04. Docker 17.06.2-ce is used to deploy and manage the HyperLedger Fabric components.

Fig. 5 illustrates the detailed configuration of the BT network. Fig. 5(a) is a conceptual diagram, and Fig. 5(b) is the deployed network. The network is composed of a single channel, two organizations, and a Raft ordering cluster [57], which consists of five orderers. Peers and orderers are annotated with red and blue rectangles in Fig. 5(b), respectively. These two organizations are identified as Org1 and Org2, respectively, and are used to represent the participating entities in the collaboration. Org1 comprises two peers, namely Peer 1 and Peer 2, while Org2 consists of a single peer, Peer 3. To ensure authority control, six certificate authorities are employed (not depicted in Fig. 5). Org1 and Org2 have one certificate authority (CA) each for distributing certificates and private keys for identity authentication, as well as one transport layer security CA for distributing public keys and private keys to enable secure communication. Those CAs are marked with pink rectangles in Fig. 5(b).

4.2. Machine learning acquisition

The operation logic in the indoor construction progress monitoring case can be formalized into three parts. The input is a set of images

Ω

, corresponding poses P, and BIM model

X_{B I M}

, which is shown in Figs. 6(a)–(c). The “MLProcess” in Table 2 is the as-is status retrieval step which aims to use an ML-based semantic segmentation method to retrieve pixel-level semantic label of as-is scene as progress information. A semantic model usually includes an encoder

F_{e n c}

to receive images as input to get the high dimensional embedding, and a decoder

F_{d e c}

to predict the semantic labels

{\hat{y}}_{i s}

of image

ω

captured from as-is scene, which is shown in Fig. 6(d).

(1)

{\hat{y}}_{i s} = F_{d e c} (F_{e n c} (ω)), ω \in Ω

“QueryFromDLT” refers to the as-planned status retrieval step, which aims to obtain the projection

y_{p l a n n e d}

from the BIM model

X_{B I M}

at the same pose p as as-is and use its semantic information as as-planned data. The

y_{p l a n n e d}

is prepared in advance and stored in the distributed ledger.

(2)

y_{p l a n n e d} = p ∙ X_{B I M}, p \in P, y_{p l a n n e d} \in Y_{p l a n n e d}

“Compare” refers to the progress estimation step comparing the as-is and as-planned status. The estimation result of i-th object in the scene

R_{p r o g r e s s}^{i}

is obtained by calculating the ratio of pixels belonging to i-th object in as-planned status

Y_{p l a n n e d}^{i}

to the pixels in the same position in as-is status with the same semantic label

{\hat{Y_{is}}}^{\hat{i}}

(3)

R_{p r o g r e s s}^{i} = \frac{{\hat{Y_{is}}}^{\hat{i}}}{Y_{p l a n n e d}^{i}}, \hat{i} = i n d e x (Y_{p l a n n e d}^{i} && Y_{p l a n n e d}^{i} = Y_{i s}^{i})

In the ML model selection process, the ML-based semantic segmentation model known as DeepLab-V3 [58] is chosen as the semantic information extractor for retrieving semantic information from vision data. The MobileNet-V3 [59] is selected as the backbone network for this ML model. After 100 epochs of training, the ML model converges. The progress estimation result is obtained in the form of a semantic point cloud in Fig. 6(e), and is visualized within the BIM model in Fig. 6(f). The colors in Figs. 6(b)–(e) correspond to specific semantic labels. In Fig. 6(f), three different statuses are represented by distinct colors: green for installed, orange for in progress, and red for uninstalled.

4.3. Machine learning conversion

To enable the ML-based progress estimation computing logic to function within the JavaScript-based BSC runtime environment, a transfer from the Python-based development environment is required. Initially trained using the Pytorch framework, the ML model is saved in the .pth format. To facilitate its usage in the BSC, it is necessary to convert the model into an intermediate protocol buffer file. Then the conversion is achieved through the utilization of the TensorFlow.js converter.

The conversion process is depicted in Fig. 7. In the first part of Fig. 7, the .pth file solely contains the variable weights of the ML model. Moving to the second part of Fig. 7, the intermediate protocol buffer file not only stores the variable weights but also includes the ML model architecture. In the third part of Fig. 7, the converted ML model consists of several components. Firstly, the model.json file saves the model architecture in JSON format. Additionally, there are four “bin” files that store the weights or parameters of the ML model. To optimize network transferring and caching performance, the variable weights are sliced into a group of smaller files, with each file being approximately 4 megabytes (MB) in size. This approach allows for enhanced performance when transferring and caching the weights.

4.4. Machine learning safe loading

Following the conversion process, a cryptographic hashing function called MD5 [60] is employed to generate a hash value for the converted ML model. This hash value is presented in Table 3. In the next step, the ML model files, along with their respective hash values, are loaded by the BSC within the BT network. By utilizing Algorithm 1 as outlined in Table 1, the BSC can verify if the ML model has been tampered with by comparing the hash values. Once the verification is complete, the BSC proceeds to load the vision data or record required for ML-based progress estimation.

4.5. Consensus-based computing

In the BSC runtime environment, ML-based progress estimation computing logic is executed as BSC based on node.js 8.9.0. The BSC is powered by fabric-shim 1.4.4 to support Fabric BSC development, and TFJS-node 3.1.0 to support ML computing on BSC. Additionally, a series of SDKs is implemented to provide a set of efficient and straightforward functions that enable the organization client to interact with the BT network based on HyperLedger Fabric. These functions are implemented using the Node.js fabric-client SDK based on node.js 13.7.0.

Table 4 shows an example input that is about to be submitted to BSC for consensus-based computing. The record is formatted as a JSON string, which includes the vision data, the uploader information, the task description, and other metadata. During endorsement, the nodes required by endorsement policy should provide their certificate issued by CA. The signed input data is then packaged as a transaction proposal to execute on the three peers simultaneously. After the progress estimation shown in Algorithm 2 in Table 2, the ML-based progress estimation result is updated into the record and stored in the distributed ledger as an immutable record. The execution time and performance are recorded for evaluation and comparison.

5. Comparison and evaluation

In this section, ablation experiments are first conducted to assess the impact of various on/off-chain configurations on the effectiveness and feasibility of deployment in real-world engineering computing tasks. The comparison and evaluation focus on accuracy, efficiency, and security. All evaluations employ quantitative measures. For accuracy assessment, the mean intersection over union (MIoU) is used as an evaluation metric, given that the actual engineering computing task in this research involves semantic segmentation. For efficiency evaluation, computing time serves as the metric. For security assessment, six attack scenarios are designed, and the ability of different methods to defend against these attacks is used as the evaluation metric.

The ablation experiment configuration is introduced in Section 5.1, while the evaluations of security, accuracy, and efficiency are comprehensively detailed in 5.2 Security evaluation, 5.3 Accuracy evaluation, 5.4 Efficiency evaluation. Furthermore, a comparative analysis between the proposed MLOB framework and two recent ML–BT integrated approaches is presented in Section 5.5.

5.1. Ablation experiments

Ablation experiments are devised to assess the impact of different on/off-chain configurations on efficiency, accuracy, and security. Three baselines are established for comparison, and their configurations are presented in Table 5. In particular, baseline-1 involves off-chain data input, baseline-2 involves off-chain ML computing with on-chain data input and result output [34], and baseline-3 pertains to off-chain result output. To ensure a fair comparison, the remaining components of the baselines remain consistent with the MLOB framework.

5.2. Security evaluation

To evaluate the security, both logic corruption and data tampering are conducted. Given the distinct architectures of the baseline and MLOB frameworks, different attack scenarios are designed to simulate different form of data tampering and logic corruption, which are summarized on Table 6. By analyzing the response to these attack scenarios, the security performance can be assessed.

Attack scenario #1

Attack scenario #1 aims to directly modify the ML computing logic in the baseline-2 method. In a local ML computing environment, the normal logic can be easily tampered with to introduce malicious logic. Obviously, the baseline-2 continues to operate without generating any errors or warnings during the execution of this malicious logic. Consequently, the baseline-2 framework proves ineffective in defending against attack scenario #1.

Attack scenario #2

Attack scenario #2 involves directly modifying the BSC, with Org2 serving as the malicious organization. As illustrated in Fig. 8, the attack scenario is conducted on BSC version 1.3, where Peer3 in Org2 modifies their hosted BSC content to include malicious logic. The invoking result of this modification is shown in Fig. 9. The transaction proposal must be verified by the blockchain network. However, since the invoking result of Peer3 in Org2 is different from that of the peers in Org1, the transaction response fails to comply with the consensus strategy and specified verification rules. As a result, the transaction proposal is considered to have failed, and the MLOB framework successfully defends against attack scenario #2.

Attack scenario #3

Attack scenario #3 involves the application of malicious logic through the upgrade of the BSC. In this scenario, Org2 installs malicious logic on its own node, Peer3, and sends a proposal to the Orderer to request an upgrade of the BSC. The result of the upgrade is depicted in Fig. 8. Since Org2 can only operate peers within its organization and cannot interfere with Org1, the maliciously updated BSC version 1.4 is only installed on Peer 3. Other peers continue to maintain BSC version 1.3, while the blockchain network reflects the latest BSC version as 1.4. After the upgrade, the BSC invoking result is shown in Fig. 10. Peers in Org1 unable to invoke the BSC with the latest version 1.4 therefore report an error. In this way, MLOB successfully defends against attacks in attack scenario #3.

Attack scenario #4

Scenario #4 of the attack involves the direct modification of data input for baseline-1. As baseline-1 solely relies on a database to store the data input, it becomes susceptible to tampering. The occurrence of such data tampering proves to be beyond the capability of baseline-1 to handle. Consequently, the ineffectiveness of baseline-1 in defending against attack scenario #4 is established.

Attack scenario #5

Scenario #5 of the attack entails the direct modification of the result output for baseline-3. In the case of baseline-3, the data input is uploaded to a blockchain network, ensuring its immutability and traceability. However, the result output is stored in an off-chain database, making it vulnerable to undetected modifications. Baseline-3 lacks the capability to detect such alterations in the result output. Although it is possible to recover the result output by invoking the process of ML computing, this endeavor demands additional effort and computational power, ultimately undermining the original purpose of a combined ML–BT system. Consequently, baseline-3 is deemed ineffective in defending against attack scenario #5

Attack scenario #6

Scenario #6 revolves around the modification of on-chain data within the MLOB system. To facilitate the blockchain network of MLOB, a CouchDB [61] is employed. This database configuration offers a graphical interface, enabling users to inspect the data stored within MLOB. Should a user attempt to tamper with the data input or result output, it becomes impossible to store or replace the original record, as illustrated in Fig. 11. Consequently, MLOB effectively safeguard against data tampering.

5.3. Accuracy evaluation

Table 7 lists the intersection over union (IoU) metric for each category. The MIoU, which averages across all categories, is used to represent overall performance. This can be calculated using Eqs. (4), (5), where

ε_{ij}

represents the number of pixels with a ground truth category i that are assigned to category j, and k represents the total number of categories. During ML acquisition, all training configurations are same, and all random seeds are fixed as a same value during ablation experiments. Therefore, for the methods that put the ML computing process on-chain, that is, MLOB, baseline-1, and baseline-3, the accuracy evaluation results are same. In Table 7, the third row shows the IoUs calculated by baseline-2, while fourth and fifth rows display the IoUs calculated from the MLOB and baseline-1 & baseline-3, respectively. The sixth row represents the difference in IoU evaluation between the results obtained from the baseline-2 and MLOB frameworks. The MIoU for the baseline-2 and MLOB frameworks is 0.770 and 0.769, respectively. The overall MIoU difference is 0.001. The semantic segmentation result of a single image is depicted in Fig. 12. In Fig. 12(c), the white portion represents the result obtained from Figs. 12(a) and (b) has the same prediction result.

(4)

{I o U}_{i} = \frac{ε_{ii}}{\sum_{j = 0}^{k} ε_{ij} + \sum_{j = 0}^{k} ε_{ji} - ε_{ii}}

(5)

M I o U = \frac{1}{k} \sum_{i = 0}^{k} {I o U}_{i}

As the ML computing model used in the MLOB framework is converted from the one in the baseline-2, accuracy evaluation results from the baseline-2 and MLOB frameworks should theoretically be identical. We speculate that the accuracy gap is due to variations in the runtime environment. Certain low-level operators, for instance, the resize operator [62], may be implemented differently in various operating environments. Although these disparities are challenging to eliminate, their impact on actual task performance is minimal. Consequently, it can be concluded that the MLOB framework achieves accuracy performance comparable to the baseline that conduct ML computing locally.

5.4. Efficiency evaluation

Fig. 13 presents the results of the efficiency comparison. The ML computing logic is executed separately in baselines and MLOB, with the efficiency evaluated based on the computing time, also referred to as latency. The latency for 20 images in the test set is recorded as frequency histogram in Fig. 13, the horizonal axis refers to the computing time range in milliseconds, the vertical axis refers to the corresponding frequency. For the baseline-2, the ML computing process is conducted on a local computer. The experiment is repeated three times, and the average computing time is recorded. For MLOB, baseline-1, and baseline-3, the ML computing process takes place on the BSC with a JavaScript runtime environment. The computing time of three peers in the BT network is recorded, and the average value is reported in Fig. 13.

The average total latency for MLOB, baseline-1, baseline-2, and baseline-3 is 0.400, 0.410, 0.169, and 0.389 s, respectively. Baseline-2 exhibits the shortest time due to its local execution of the ML computing process. In contrast, MLOB requires an additional 0.231 s compared to baseline-2 due to the efficiency variance between the BSC runtime environment and the local runtime environment. This time is necessary for MLOB to verify the legality of transaction content and signatures. In comparison, baseline-1 requires a longer time as it needs to load data from the off-chain world for computing. On the other hand, baseline-3 exhibits a shorter time as it omits the necessity of storing the result output into the blockchain network.

The ablation experiment shows the necessity of putting both data input, ML computing, and result output on blockchain. The evaluation results reveal that the MLOB framework can significantly enhance security while maintaining accuracy, albeit with a slight compromise in efficiency. As for accuracy evaluation, the MIoU gap between MLOB and the best result from baseline-2 is a mere 0.001, which is negligible for most engineering computing tasks. Regarding efficiency, the latency difference is at the millisecond level. Specifically, for a single calculation task, the MLOB framework is 0.231 s slower than the best result from baseline-2. According to research conducted by Zheng et al. [63], when the delay is less than 0.4 s, ML-based engineering computing can efficiently adapt to collaborative work on engineering computing tasks. In terms of security evaluation, the MLOB framework effectively defends against various forms of logical corruption and data tampering, whereas baseline methods fall short in this aspect.

5.5. Comparison with existing solutions

In the performance comparison, two recent integrated ML–BT methods are utilized. The first method, TFL-DT [64], introduces a trust evaluation scheme for federated learning. This scheme is integrated with a BT-based federated learning framework called FedSeg [65] to during the comparison. Similar as the deployment of MLOB, three nodes are configured for joint training. Subsequently, solely the two most reliable local models are selected for global aggregation in each training round. Other configurations and hyperparameter settings remain similar to the default configuration.

The second method used for comparison is CoBC [49], which enables each participant to maintain a customized local ML model. It employs a novel collaborative inference result aggregation strategy to obtain the output. During the comparison, three nodes are involved in the local model training and collaborative inference. Each organization maintains two model variants using the Monte-Carlo dropout technique. Other configurations and hyperparameter settings remain similar to MLOB.

Table 8 presents the IoUs for each category of the MLOB, TFL-DT, and CoBC. Owing to substantial disparities in framework architecture, design objectives, and implementation particulars, the security performance of these methods varies considerably. Consequently, an analysis of accuracy, efficiency, and security is conducted through both qualitative and quantitative means.

The fourth row of Table 8 displays the IoUs for each category of TFL-DT. With regard to accuracy performance, the MIoU of MLOB and TFL-DT are 0.77 and 0.83, respectively. In terms of efficiency performance, the computing time for MLOB and TFL-DT are 0.400 and 0.182 s, respectively. For security performance, TFL-DT presents an innovative trust evaluation scheme to enhance training reliability and stability. During the training stage, the joint training process results in superior accuracy for TFL-DT, and the trust evaluation scheme ensures security. However, TFL-DT fails to consider the inference stage, which conducts the process locally and may expose it to potential security threats, such as logic corruption, leading to failure modes similar to attack scenario #1 in Section 5.2.

The fifth row of Table 8 illustrates the IoUs for each category of CoBC. Concerning accuracy performance, the MIoU of MLOB and CoBC are 0.77 and 0.75, respectively. In terms of efficiency comparison, the computing time for MLOB and CoBC are 0.400 and 1.023 s, respectively. Regarding security performance, the CoBC framework adopts a local training and collaborative inference strategy for both training and inference stages. During the training stage, the local training strategy significantly reduces the risk of privacy leakage and ensures data security, but may lead to a loss of accuracy. During the inference stage, the collaborative inference strategy ensures computational security but increases redundancy and reduces efficiency.

To summarize, TFL-DT excels in accuracy, efficiency, and training security but overlooks security considerations when implementing the trained ML model in engineering computing tasks. CoBC offers substantial security advantages, but at the cost of compromised accuracy and efficiency. On the other hand, MLOB provides a secure environment for engineering computing while making a modest compromise in terms of accuracy and efficiency.

6. Discussion

Unlike existing solutions that primarily focus on data security, the MLOB framework introduced in this study ensures the security of both data and computational processes by incorporating them into the blockchain and executing them as BSCs. As BT requires inherent redundancy for consensus and transaction validation, the enhanced security in MLOB does entail a trade-off between efficiency and accuracy due to the occupied computing resources. An evaluation comparing with baseline methods indicates that this rebalancing maintains satisfactory efficiency and accuracy levels, enabling MLOB to effectively carry out engineering computing tasks within a highly secure requirement.

To gain a comprehensive understanding of the advantages and disadvantages of the MLOB framework, this study compares it with similar concepts, such as edge computing and federated learning. The fundamental concept of the edge computing model is to shift the data processing, storage, and computing operations, originally reliant on the centralized cloud, to the network edge in close proximity to terminal devices [66]. While edge computing primarily emphasizes safeguarding the data and computing security of participants, this MLOB paradigm focuses on ensuring the overall security of the engineering computing system, preventing participants from tampering with engineering calculation data and logic. Moreover, edge computing-based systems typically exhibit lower latency and enhanced computational efficiency. Consequently, when users prioritize computing efficiency and are willing to compromise some security for it, the edge computing paradigm is more suitable.

Federated learning is a distributed ML framework where individual local users need not upload all their raw data to a central server for training. Instead, they train their local models using privacy-related data, and these local models are then aggregated into a global model on the central server [67]. The primary focus of federated learning is to safeguard the ML acquisition process. It effectively prevents attackers from accessing and tampering with training data, as well as compromising, attacking, or destroying the ML models.

On the other hand, the MLOB paradigm concentrates on protecting the entire process of applying ML to specific engineering computing tasks. MLOB and federated learning paradigms are not contradictory but complementary with each other. If users aim to aggregate data from multiple participants to train a customized model with high accuracy, while accepting a trade-off in computational efficiency, the federated learning paradigm can be optionally introduced during the ML acquisition stage.

This research investigates a novel ML-BT integrated solution to ensure both data and computational security in engineering computing. While this approach has numerous advantages, not all engineering data require blockchain security, and not all computing logic needs to be executed on the blockchain. An actual engineering computing task is typically complex, involving multiple workflows and data types. According to existing best practices, key data in engineering computing tasks, including original data input, critical intermediate data, or result output, are suitable for storage on the blockchain [34]. However, computing logic that processes highly private personal information, handles dynamic data, or deals with unstructured large files is generally not suitable for execution on BSC due to internal accessibility control, processing efficiency, and data redundancy concerns [68].

The MLOB framework does encounter certain challenges and limitations, as outlined below:

(1) Limited support for latency-sensitive scenarios. With the growing demand for real-time analysis in engineering computing tasks, the MLOB framework falls short in providing adequate support for low-latency requirements. While security improvements are crucial, they should not come at the expense of efficiency for tasks that heavily rely on real-time capabilities.

(2) Lack of a user-friendly interface. Although the prototype developed in this study is ready for deployment, it can only be operated through the command line. This limitation increases the complexity of understanding its functionalities and hampers efficient interaction. Consequently, it diminishes productivity and overall effectiveness of the application.

The managerial implications of this study are substantial. Engineering managers and policy makers can benefit from both technical innovation and economic advancements. From a technical innovation perspective, MLOB encourages organizations and enterprises to foster innovation in engineering practices by integrating advanced technologies, including ML and BT. Ultimately, this can lead to more competitive engineering operations, increased productivity, and the attraction of talent interested in cutting-edge technologies. Economically, organizations can mitigate risks associated with data tampering, logic corruption, and operational inefficiencies by enhancing computational security. This approach fosters economic resilience, cost savings, and optimized resource allocation. Furthermore, the ability to maintain security while streamlining operations may enhance overall productivity and competitiveness in the market.

To assist potential users in translating theoretical insights into practical applications, a suggested action plan is provided. This plan includes pilot exploration, a training program, performance monitoring, and iterative improvement, detailed as follows:

(1) Pilot exploration: Encourage collaboration among information technology (IT), engineering, and cybersecurity teams to ensure comprehensive implementation of the MLOB framework, assessing its functionality and compatibility through small-scale initiatives.

(2) Training program: Develop comprehensive training sessions for staff to familiarize them with the MLOB framework, focusing on its features, benefits, and security protocols to enhance user proficiency.

(3) Performance monitoring: Establish metrics and benchmarks to continuously evaluate the MLOB system’s performance, tracking accuracy, efficiency, and security outcomes to ensure operational goals are met.

(4) Iterative improvement: Use insights gained from pilot exploration and performance monitoring to refine the framework and processes, promoting an ongoing cycle of enhancement based on user feedback and evolving needs.

7. Conclusions

Despite the growing use of ML in engineering computing, existing research on integrating ML with BT tends to focus solely on data security, neglecting the importance of ML computational security. To address the potential risks of logic corruption, this research proposes a practical framework that can ensure both the data and computational security in engineering computing. The findings indicate that the MLOB framework can significantly enhance the computational security of engineering processes without increasing computing power requirements.

In more detail, with the same hardware configuration, the proposed method successfully defends against all attack scenarios, outperforming all other baseline methods. The accuracy performance of the proposed method, measured as 0.769 MIoU, is slightly lower than the 0.770 MIoU achieved by the best baseline method. As for efficiency, the proposed method takes 0.400 s for a single engineering task, resulting in a latency increase of 0.231 s compared to the most efficient baseline method. In certain engineering computing tasks, the priority may be given to security over accuracy and efficiency, which justifies this trade-off.

This research possesses both practical and theoretical significance. From a practical perspective, the proposed MLOB paradigm ensures that data input, ML computing processes, and result output are all conducted on the blockchain, guaranteeing decentralization, immutability, and traceability. A real-world indoor construction progress monitoring task is employed for validation, demonstrating that the security, accuracy, and efficiency of the MLOB framework meet the actual requirements of industrial practice. This surprising discovery alleviate doubts regarding the limitations of computing resources in practice, instilling confidence in users to integrate it into their workflows.

From a theoretical standpoint, this study utilizes a “balance triangle” to illustrate the trade-offs between security, efficiency, and accuracy in the ML–BT integration process, conducting a performance evaluation and comparison based on these three aspects. This representation highlights a research gap: the insufficient consideration of computational security. The research objective is to propose a novel paradigm, MLOB, to enhance security while maintaining accuracy and efficiency. The “triangle” representation provides a scientific framework for understanding the trade-off among accuracy, efficiency, and security, serving as a foundation for future research questions and objectives.

To enhance the efficacy and user-friendliness of MLOB, future efforts should be directed towards expanding the platform. Firstly, the primary focus of future developments will be on optimizing the efficiency of MLOB, with the objective of enhancing the “triangular balance” in a spiral manner so that MLOB can accommodate a wider range of engineering computing tasks, particularly those that are latency sensitive. Secondly, a user interface will be designed to elevate the application's level of usability, accessibility, and overall user satisfaction. These proposed enhancements will serve to facilitate the integration of ML and BT, which will encourage engineers to adopt smart technologies in their computing workflows, and to contribute to digital transformation across the industry over the long term.

CRediT authorship contribution statement

Zhiming Dong: Writing – review & editing, Writing – original draft, Validation, Methodology, Conceptualization. Weisheng Lu: Writing – review & editing, Supervision, Resources, Funding acquisition.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

References

Publishing order | Descend order by publishing year | Descend order by cited within

[1]	Buede DM, Miller WD.The engineering design of systems: models and methods.John Wiley & Sons, Hoboken (2016)

[2]	Sinha R, Paredis CJ, Liang VC, Khosla PK.Modeling and simulation methods for design of engineering systems.J Comput Inf Sci Eng 2001; 1(1):84-91.

[3]	Nagrath I, Gopal M.Control systems engineering.New Age International, Hong Kong (2006)

[4]	Sioshansi R, Conejo AJ.Optimization in engineering.Springer International Publishing, Cham (2017)

[5]	Willard J, Jia X, Xu S, Steinbach M, Kumar V.Integrating scientific knowledge with machine learning for engineering and environmental systems.ACM Comput Surv 2023; 55(4):1-37.

[6]	Penumuru DP, Muthuswamy S, Karumbu P.Identification and classification of materials using machine vision and machine learning in the context of Industry 4.0.J Intell Manuf 2020; 31(5):1229-1241.

[7]	Frank M, Drikakis D, Charissis V.Machine-learning methods for computational science and engineering.Computation 2020; 8(1):15.

[8]	Masana M, Liu X, Twardowski B, Menta M, Bagdanov AD, Van J De Weijer.Class-incremental learning: survey and performance evaluation on image classification.IEEE Trans Pattern Anal Mach Intell 2023; 45(5):5513-5533.

[9]	Jeong D.Artificial intelligence security threat, crime, and forensics: taxonomy and open issues.IEEE Access 2020; 8:184560-184574.

[10]	Nadeem A, Vos D, Cao C, Pajola L, Dieck S, Baumgartner R, et al.SoK: explainable machine learning for computer security applications.In: Proceedings of the 2023 IEEE 8th European Symposium on Security and Privacy (EuroS&P); 2023 Jul 3–7; Delft, Netherlands. New York City: IEEE; 2023. p. 221–40.

[11]	Qadir S, Noor B.Applications of machine learning in digital forensics.In: Proceedings of the 2021 International Conference on Digital Futures and Transformative Technologies (ICoD T2); 2021 May 20–21; Islamabad, Pakistan. New York City: IEE E; 2021. p. 1–8.

[12]	Newhart KB, Marks CA, Rauch-Williams T, Cath TY, Hering AS.Hybrid statistical-machine learning ammonia forecasting in continuous activated sludge treatment for improved process control.J Water Process Eng 2020; 37:101389.

[13]	Lu W, Lou J, Wu L.Combining smart construction objects-enabled blockchain oracles and signature techniques to ensure information authentication and integrity in construction.J Comput Civ Eng 2023; 37(6):04023031.

[14]	Venkatesan C, Karthigaikumar P, Varatharajan R.A novel LMS algorithm for ECG signal preprocessing and KNN classifier based abnormality detection.Multimedia Tools Appl 2018; 77(8):10365-10374.

[15]	Lu W, Wu L, Zhao R, Li X, Xue F.Blockchain technology for governmental supervision of construction work: learning from digital currency electronic payment systems.J Constr Eng Manage 2021; 147(10):04021122.

[16]	Lu W, Wu L, Zhao R.Rebuilding trust in the construction industry: a blockchain-based deployment framework.Int J Constr Manag 2023; 23(8):1405-1416.

[17]	Papernot N, McDaniel P, Sinha A, Wellman MP.SoK: security and privacy in machine learning.In: Proceedings of the 2018 IEEE European Symposium on Security and Privacy (EuroS&P); 2018 Apr 24–26; London, UK. New York City: IEEE; 2018. p. 399–414.

[18]	Currie R.Software engineer accused of stealing $300k from employer was ‘inspired by office space’ [Internet].San Francisco: Situation Publishing; 2023 Jan 13 [cited 2023 Feb 6]. Available from: https://www.theregister.com/2023/01/13/office_space_zulily_theft/?td=rt-3a.

[19]	Fang M, Cao X, Jia J, Gong NZ.Local model poisoning attacks to byzantine-robust federated learning.In: Proceedings of the 29th USENIX Conference on Security Symposium; 2020 Aug 12–14; Berkeley, C A, US A. Berkeley: The USENIX Association; 2020. p. 1623–40.

[20]	Lee JH, Shin J, Realff MJ.Machine learning: overview of the recent progresses and implications for the process systems engineering field.Comput Chem Eng 2018; 114:111-121.

[21]	Guo J, Gao H, Liu Z, Huang F, Zhang J, Li X, et al.ICRA: an intelligent clustering routing approach for UAV ad hoc networks.IEEE Trans Intell Transp Syst 2023; 24(2):2447-2460.

[22]	Yuan L, Lu W, Xue F, Li M.Building feature‐based machine learning regression to quantify urban material stocks: a Hong Kong study.J Ind Ecol 2023; 27(1):336-349.

[23]	Yuan L, Lu W, Wu Y.Characterizing the spatiotemporal evolution of building material stock in China’s greater bay area: a statistical regression method.J Ind Ecol 2023; 27(6):1553-1566.

[24]	Guo J, Li X, Liu Z, Ma J, Yang C, Zhang J, et al.Trove: a context-awareness trust model for vanets using reinforcement learning.IEEE Internet Things J 2020; 7(7):6647-6662.

[25]	Dong Z, Chen J, Lu W.Computer vision to recognize construction waste compositions: a novel boundary-aware transformer (BAT) model.J Environ Manage 2022; 305:114405.

[26]	Dong Z, Wang J, Cui B, Wang D, Wang X.Patch-based weakly supervised semantic segmentation network for crack detection.Constr Build Mater 2020; 258:120291.

[27]	Parashar A, Parashar A, Shabaz M, Gupta D, Sahu AK, Khan MA.Advancements in artificial intelligence for biometrics: a deep dive into model-based gait recognition techniques.Eng Appl Artif Intell 2024; 130:107712.

[28]	Asmitha P, Rupa C, Nikitha S, Hemalatha J, Sahu AK.Improved multiview biometric object detection for anti spoofing frauds.Multimedia Tools Appl 2024; 83(33):80161.

[29]	Teoh YK, Gill SS, Parlikad AK.Iot and fog computing based predictive maintenance model for effective asset management in Industry 4.0 using machine learning.IEEE Internet Things J 2021; 10(3):2087-2094.

[30]	L Găzăroiu, Andronie M, Iatagan M, Geam Mănu, Dijm Iărescu.Deep learning-assisted smart process planning, robotic wireless sensor networks, and geospatial big data management algorithms in the internet of manufacturing things.ISPRS Int J Geoinf 2022; 11(5):277.

[31]	Li M, Xue F, Wu Y, Yeh AG.A room with a view: automatic assessment of window views for high-rise high-density areas using city information models and deep transfer learning.Landsc Urban Plan 2022; 226:104505.

[32]	Chan IY, Liu AM.Effects of neighborhood building density, height, greenspace, and cleanliness on indoor environment and health of building occupants.Build Environ 2018; 145:213-222.

[33]	Leung M, Wang C, Chan IY.A qualitative and quantitative investigation of effects of indoor built environment for people with dementia in care and attention homes.Build Environ 2019; 157:89-100.

[34]	Lu W, Wu L, Xu J, Lou J.Construction E-inspection 2.0 in the COVID-19 pandemic era: a blockchain-based technical solution.J Manage Eng 2022; 38(4):04022032.

[35]	Solanki S, Solanki AD.Review of deployment of machine learning in blockchain methodology.Int Res J Adv Sci Hub 2020; 2(9):14-20.

[36]	Shinde R, Nilakhe O, Pondkule P, Karche D, Shendage P.Enhanced road construction process with machine learning and blockchain technology.In: Proceedings of the 2020 International Conference on Industry 4.0 Technology (I4Tech); 2020 Feb 13–15, Pune, India. New York City: IEE E; 2020. p. 207–10.

[37]	Wong P, Chia F, Kiu M, Lou E.The potential of integrating blockchain technology into smart sustainable city development.In: Proceedings of the IOP Conference Series: Earth and Environmental Science; 2020 Sep 18–20; Changsha, China. Bristol: IOP Publishing; 2020. p. 012020.

[38]	Xu J, Lu W, Wu L, Lou J, Li X.Balancing privacy and occupational safety and health in construction: a blockchain-enabled P-OSH deployment framework.Saf Sci 2022; 154:105860.

[39]	Rk K, Kallapu B, Dodmane R, Thota S, Sahu AK.Enhancing cloud communication security: a blockchain-powered framework with attribute-aware encryption.Electronics (Basel) 2023; 12(18):3890.

[40]	Podder D, Deb S, Banik D, Kar N, Sahu AK.Robust medical and color image cryptosystem using array index and chaotic S-box.Cluster Comput 2024; 27(4):4321.

[41]	Jovanovic Z, Hou Z, Biswas K, Muthukkumarasamy V.Robust integration of blockchain and explainable federated learning for automated credit scoring.Comput Netw 2024; 243:110303.

[42]	Li X, Zeng J, Chen C, Chi H, Shen GQ.Smart work package learning for decentralized fatigue monitoring through facial images.Comput Aided Civ Infrastruct Eng 2023; 38(6):799-817.

[43]	Kasyap H, Tripathy S.Privacy-preserving and byzantine-robust federated learning framework using permissioned blockchain.Expert Syst Appl 2024; 238:122210.

[44]	Wu L, Lu W, Chen C.Strengths and weaknesses of client-server and peer-to-peer network models in construction projects.Int J Constr Manag 2023; 24(12):1349-1363.

[45]	Peethambaran G, Naikodi C, Suresh L.An ensemble learning approach for privacy–quality–efficiency trade-off in data analytics.In: Proceedings of the 2020 International Conference on Smart Electronics and Communication (ICOSE C); 2020 Sep 10–12; Trichy, India. New York City: IEE E; 2020. p. 228–35.

[46]	Cooper AF, Levy K, De Sa C.Regulating accuracy-efficiency trade-offs in distributed machine learning systems.2020. SSR N 3650497.

[47]	Al-Marridi AZ, Mohamed A, Erbad A.Reinforcement learning approaches for efficient and secure blockchain-powered smart health systems.Comput Netw 2021; 197:108279.

[48]	Jiang R, Li J, Bu W, Shen X.A blockchain-based trustworthy model evaluation framework for deep learning and its application in moving object segmentation.Sensors 2023; 23(14):6492.

[49]	Feng X, Li L, Wang T, Xu W, Zhang J, Wei B, et al.CoBC: a blockchain-based collaborative inference system for the internet of things.IEEE Internet Things J 2023; 10(24):21389-21400.

[50]	Wang SJ, Pei K, Yang J.SmartInv: multimodal learning for smart contract invariant inference.In: Proceedings of the 2024 IEEE Symposium on Security and Privacy (S P); 2024 May 19–23; San Francisco, C A, US A. New York City: IEE E; 2024.

[51]	Adel K, Elhakeem A, Marzouk M.Decentralizing construction ai applications using blockchain technology.Expert Syst Appl 2022; 194:116548.

[52]

Amershi S, Begel A, Bird C, DeLine R, Gall H, Kamar E, et al.Software engineering for machine learning: a case study.In: Proceedings of the 2019 IEE E/AC M 41st International Conference on Software Engineering: Software Engineering in Practice (ICSE-SEI P); 2019 May 25–31; Montreal, Q C, Canada. New York City: IEE E; 2019. p. 291–300.

[53]	Bertolini M, Mezzogori D, Neroni M, Zammori F.Machine learning for industrial applications: a comprehensive literature review.Expert Syst Appl 2021; 175:114820.

[54]	Kirillov A, Mintun E, Ravi N, Mao H, Rolland C, Gustafson L, et al.Segment anything.2023. arXiv: 2304.02643.

[55]	Gerard C.Practical machine learning in JavaScript: TensorFlow.js for web developers.Springer, Berlin 2021; 25-43.

[56]	Wu H, Zhong B, Li H, Guo J, Wang Y.On-site construction quality inspection using blockchain and smart contracts.J Manage Eng 2021; 37(6):04021065.

[57]	Ongaro D, Ousterhout J.In search of an understandable consensus algorithm.In: Proceedings of the 2014 USENIX Annual Technical Conference (USENIX ATC 14); 2014 Jun 19–20; Philadelphia, P A, US A. Berkeley: The USENIX Association; 2014. p. 305–19.

[58]	Chen LC, Papandreou G, Schroff F, Adam H.Rethinking atrous convolution for semantic image segmentation.2017. arXiv: 1706.05587.

[59]	Howard A, Sandler M, Chu G, Chen LC, Chen B, Tan M, et al.Searching for mobilenetv3.In: Proceedings of the IEEE/CVF International Conference on Computer Vision; 2019 Oct 27–Nov 2; Seoul, Republic of Korea. New York City: IEE E; 2019. p. 1314–24.

[60]	Dobbertin H.Cryptanalysis of MD5 compress.In: Proceedings of the EUROCRYPT’ 96 Rump Session; 1996 May 12–16; Zaragoza, Spain. Bellevue: International Association for Cryptologic Research; 1996. p. 9671–82.

[61]	Anderson JC, Lehnardt J, Slater N.CouchDB: the definitive guide: time to relax.editors. Sebastopol: O’Reilly Media; 2010.

[62]	Savsunenko O.How tensorflow’s tf.image. resize stole 60 days of my life [Internet]. Edwards: HackerNoon; 2018 Jan 23 [cited 2023 Dec 14]. Available from: https://hackernoon.com/how-tensorflows-tf-image-resize-stole-60-days-of-my-life-aba5eb093f35.

[63]	Zheng P, Li S, Xia L, Wang L, Nassehi A.A visual reasoning-based approach for mutual-cognitive human-robot collaboration.CIRP Ann 2022; 71(1):377-380.

[64]	Guo J, Liu Z, Tian S, Huang F, Li J, Li X, et al.TFL-DT: a trust evaluation scheme for federated learning in digital twin for mobile networks.IEEE J Sel Areas Comm 2023; 41(11):3548-3560.

[65]	Miao J, Yang Z, Fan L, Yang Y.FedSeg: class-heterogeneous federated learning for semantic segmentation.In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition;2023 Jun 17–24; Vancouver, B C, Canada. New York City: IEE E; 2023. p. 8042–52.

[66]	Hua H, Li Y, Wang T, Dong N, Li W, Cao J.Edge computing with artificial intelligence: a machine learning perspective.ACM Comput Surv 2023; 55(9):1-35.

[67]	Li L, Fan Y, Tse M, Lin KY.A review of applications in federated learning.Comput Ind Eng 2020; 149:106854.

[68]	Kayikci S, Khoshgoftaar TM.Blockchain meets machine learning: a survey.J Big Data 2024; 11(1):9.

RIGHTS & PERMISSIONS

THE AUTHOR