Abstract
Construction safety monitoring increasingly uses point clouds and digital-twin representations to locate materials, machinery, and changing site conditions. These tools can improve hazard detection, but they can also expose worker presence, movement, task patterns, or proximity to high-risk zones. This manuscript presents a research protocol that treats privacy as a design constraint in safety monitoring from fixed infrastructure. The protocol synthesizes two evidence streams: construction studies showing how point-cloud segmentation and digital-twin workflows support site understanding, and transferable industrial federated-learning methods that limit raw-data sharing while addressing update leakage and poisoning risk. On that basis, the manuscript proposes an architecture in which worker-related inference is performed locally at the site edge, model updates are exchanged through federated learning, and protected layers are subject to differentiated privacy budgets, encrypted aggregation where required, and update validation before aggregation. The methodological contribution is a defined decision sequence linking zone occupancy inference, hazard scoring, privacy endpoints, and model exchange without personal wearables. The manuscript specifies variables, attack assumptions, comparators, evaluation domains, and failure cases. It also defines measurable privacy outcomes as separate endpoints for membership inference, excess occupancy inference, and trajectory reconstruction from transmitted updates or retained outputs. No empirical results are reported; the article sets out a prospective protocol for later validation in construction settings.
1. Introduction
Construction safety monitoring can use site sensing, point clouds, digital-twin representations, and machine learning to support progress review, object recognition, and hazard analysis. In this setting, the technical problem is not only whether a machine, scaffold, opening, or obstructed access route can be recognized, but also whether the same system can be deployed without turning worker movement into a continuously traceable record. That concern is specific to construction because site monitoring often uses infrastructure that observes shared workspaces, shift changes, and transient crews rather than stable room occupancy. A privacy-preserving design must therefore preserve the safety utility of spatial inference while limiting the collection, retention, and exchange of worker-related signals. [1] [2]
The closest construction-domain references show that point-cloud segmentation and digital-twin workflows can support site understanding. A BIM-guided virtual-to-real framework for construction point-cloud segmentation demonstrates that synthetic point clouds can train models for real scenes without real-scene annotation, while a digital-twin-based construction management framework uses point-cloud semantic segmentation to reconstruct site objects for safety pre-control. These studies establish feasibility for site-aware monitoring, but neither is organized around worker privacy as the primary design objective. They therefore leave open a practical question: how can one infer hazard-relevant site conditions while avoiding unnecessary exposure of worker-related information? [1] [2]
This protocol addresses that question by integrating construction sensing with transferable privacy-preserving learning methods. Industrial federated-learning studies support the use of distributed training to avoid central collection of raw data, while also showing that exchanged gradients or parameter updates can leak sensitive information and can be manipulated by malicious participants. For sensor-rich industrial environments, hierarchical privacy-budget allocation, homomorphic encryption, anomaly detection on updates, clipping, and integrity verification have been proposed as countermeasures. An occupant-centered monitoring method in public buildings is also relevant because it infers location from fixed infrastructure and timetables without wearables. In this manuscript, that case is used only as a design analogy for non-wearable monitoring; construction sites require different assumptions because site geometry, crews, and tasks change over time. [3] [4] [5] [6]
2. Research question and evidence scope
The research question is: how can a construction-site monitoring system infer and share safety-relevant spatial information while minimizing the privacy exposure of workers who are present in the monitored area? The question is narrow. It does not ask how to optimize all aspects of construction management, nor does it assume that privacy can be ensured by policy statements alone. It asks for a technically explicit protocol that permits hazard monitoring from fixed infrastructure and distributed learning while constraining the information that can be derived about worker presence, movement, or routines. [1] [2]
The evidence base is a targeted narrative synthesis rather than a systematic review. The construction layer consists of studies on point-cloud segmentation and digital-twin workflows that support semantic understanding of active sites. The privacy layer consists of transferable industrial methods for federated learning, homomorphic encryption, differential privacy, secure update exchange, and poisoning-aware validation. The occupant-centered monitoring analogy is used only to support the idea that fixed sensors can replace wearable devices when the goal is exposure or presence inference rather than personal telemetry. No claim is made that the cited methods are directly deployable unchanged in construction sites; instead, they define design options, constraints, and failure modes for the proposed protocol. [1] [2] [3] [4] [5] [6]
3. Conceptual model: what must be protected and what may be inferred
The protocol distinguishes four information classes. First, site-state variables describe legitimate safety targets: object class labels for barriers, machinery, temporary storage, access paths, openings, and hazardous zones; geometric occupancy of those zones by non-person objects; and change events over time. Second, worker occupancy variables indicate whether one or more persons are present in a zone during a short decision window. Third, worker trajectory variables describe ordered movement across zones over repeated windows and can reveal routines or task patterns. Fourth, trainable parameters and gradients encode site-state patterns but may also leak occupancy or trajectory information through inference attacks. The privacy objective is to preserve site-state inference while constraining leakage about worker occupancy, and especially worker trajectories, from stored outputs and transmitted updates. [3] [4] [6]
The proposed system therefore treats worker privacy as a bounded inference problem. The system may infer anonymized zone occupancy when that inference is necessary to trigger a safety rule, but it should not retain or exchange person-level trajectories unless a reviewed incident protocol specifically requires short-term preservation. Occupancy, presence duration, and trajectory are not treated as interchangeable constructs: occupancy is a binary or count-based state within one zone and one time window; presence duration is the time spent in that zone within a bounded retention period; trajectory is a linked sequence across windows and zones. This distinction matters because many construction safety decisions require only a local and short-lived answer, such as whether a person is inside an exclusion zone, rather than an archival record of where that person moved during the day. The protocol operationalizes that principle through local processing, short retention, and privacy-controlled model exchange. [6] [3]
4. Proposed methodology
The proposed architecture has four layers: sensing, local inference, federated update exchange, and supervisory validation. Sensing uses site-installed or temporarily mounted non-wearable infrastructure, such as fixed-position LiDAR during a monitoring interval, legally permissible fixed cameras, and environmental sensors already used for site control. No wearable device is required. The edge model performs point-cloud segmentation and zone classification on the site side, producing anonymized occupancy and hazard maps. Only model updates or bounded summary statistics leave the site. Distributed training uses federated learning so that raw point clouds and image streams remain local. Privacy controls are applied to selected model layers, with stronger perturbation on parameters judged more likely to encode worker-related spatial patterns. [1] [2] [3] [6]
Operational variables are defined as follows. Let x_t be the fixed-infrastructure observation at time t, z_t the set of geometric zones extracted from the digital twin, y_t the hazard label for each zone, and o_t the anonymized occupancy indicator for each zone. The local model f_theta maps x_t to (y_t, o_t). Privacy exposure is not treated as a single undifferentiated probability. Instead, the protocol defines three separate endpoints: r_t^(mem), the success rate of membership or sample-presence inference against transmitted updates; r_t^(occ), the success rate of inferring zone occupancy from transmitted updates or stored outputs beyond what is necessary for alerting; and r_t^(traj), the success rate of reconstructing repeated cross-window movement patterns from updates, released models, or retained outputs. A privacy event occurs when any endpoint exceeds its preregistered acceptance bound under the specified attack model. [3] [4] [6]
Default operational rule: for each hazard zone h at time t, compute occupancy probability p_t(h)=P(o_t(h)=1|x_t) and a normalized hazard severity score s_t(h) in [0,1]. The severity score is assigned by a preregistered rubric with fixed values for hazard classes and modifiers for immediacy; by default, exclusion-zone breach and moving-equipment collision risk are scored higher than housekeeping or congestion hazards, and all scores are mapped to [0,1] before deployment. The system emits a local alert if and only if p_t(h) >= tau_occ and s_t(h) >= tau_haz. If the rule is not satisfied, the system stores only aggregate zone counts and hazard summaries. To avoid post hoc tuning, tau_occ and tau_haz should be selected before validation by a documented calibration procedure on development data or synthetic scenarios; one default setting is the pair that satisfies a preregistered missed-hazard constraint while minimizing the number of retained occupancy records. Sensitivity analysis may vary each threshold over a fixed grid, but the conjunctive rule itself remains unchanged. [2] [6]
The federated-learning component is adapted from industrial privacy literature in four steps. First, layer-sensitive privacy budgeting assigns stronger protection to parameters associated with greater inferred sensitivity. A reproducible default rule is to estimate each protected layer's sensitivity from the normalized magnitude of its clipped update and to allocate the round budget proportionally, while tracking cumulative privacy loss by layer and by round. Second, each client clips its protected-layer update before privacy perturbation and, where encryption is used, encrypts the perturbed protected subset for aggregation. Third, update validation is applied to the plaintext validation features that remain available to the coordinator: directional consistency with the recent global descent direction, scale stability relative to the interquartile range of client update norms, and inter-layer similarity relative to the current round median. An update is accepted only if it passes at least two of these three checks; otherwise it is excluded from aggregation for that round and the previous global model is retained for that client's contribution. Fourth, integrity verification may be recorded through a ledger or equivalent audit mechanism if the deployment requires it. The primary comparator is standard FedAvg-style exchange without these added privacy and validation steps. [3] [4] [5]
5. Decision sequence and algorithmic workflow
The proposed workflow is sequential. Step 1: collect site observations from non-wearable infrastructure and align them to the digital-twin coordinate frame. Step 2: segment the point cloud into structural and temporary objects and derive zone occupancy maps. Step 3: compute a hazard score for each zone from a preregistered site rubric that maps hazard class and immediacy to a normalized severity value. Step 4: apply the conjunctive alert rule and generate a local warning if both occupancy and severity thresholds are exceeded. Step 5: package the model update, not the raw observation, for federated exchange. Step 6: clip protected-layer updates, add layer-aware privacy noise, and, if configured, encrypt the protected subset before transmission. Step 7: validate each received update against the acceptance rule, aggregate accepted updates, and redistribute the global model. Step 8: audit retained outputs and released models for privacy leakage and alert fidelity under the predefined attack model. [1] [2] [3] [4]
This sequence is deliberately conservative. It favors local safety action over delayed central reconstruction. The design assumes that the main task is to detect safety-relevant site states, not to archive continuous worker movement. It also assumes that the site topology can be represented adequately by the digital twin for the duration of the monitoring window. Where the site layout changes rapidly, the system must refresh the twin before inference; otherwise the location mapping may become stale and lead to false occupancy assignment. [1] [2]
Comparator definition: the primary comparator is a centralized monitoring pipeline that uploads raw point clouds or raw frames to a server for model training and safety analytics. A secondary comparator is a distributed pipeline without explicit privacy controls, as in standard federated averaging. The proposed protocol is better only if it reduces reconstructable worker information, maintains hazard detection that is operationally adequate for site use, and does not increase unsafe alert latency beyond the site’s tolerance. These are joint criteria; privacy gain is not sufficient if the system misses a critical hazard. [3] [4]
6. Prospective evaluation protocol
Because this is a research protocol, evaluation is defined prospectively. The first outcome domain is safety utility, measured by zone-level segmentation agreement, occupancy classification fidelity, alert latency, and missed-hazard rate. The second domain is privacy exposure, measured separately for the three estimands defined above: membership or sample-presence inference against updates, excess occupancy inference from outputs or updates, and trajectory reconstruction across windows. The third domain is system integrity, measured by robustness to malformed or malicious updates and stability under site-layout changes. The default attack model is an honest-but-curious coordinating server that can inspect transmitted updates and released global models, plus an external evaluator who can inspect retained outputs; both are evaluated under white-box access to model parameters, while black-box tests are reported as sensitivity analyses rather than the primary endpoint. [3] [4] [1] [2]
Validation should proceed in three phases. Phase A uses a controlled synthetic or BIM-derived environment to test whether the model can learn site semantics under the proposed privacy constraints. Phase B uses real construction point clouds from a limited set of active sites. For alert-correctness review in Phase B, events should be sampled by a preregistered stratified rule that includes both alerts and non-alert windows and oversamples occluded or rapidly changing zones. Ground truth should be established by independent dual review of the corresponding site observations and digital-twin alignment, with disagreements resolved by adjudication and reported explicitly. Phase C tests federated deployment across multiple sites or site zones with heterogeneous layouts. All phases should report failure cases, especially false occupancy in occluded regions, confusion between workers and temporary objects, and degradation caused by rapid topology change. [1] [2] [6]
The validation design should include a focused ablation sequence with one default comparator chain: centralized raw-data training, standard FedAvg-style exchange, FedAvg with clipping only, FedAvg with clipping plus differential privacy, FedAvg with clipping plus differential privacy plus homomorphic encryption on protected layers, and the full protocol with update validation added. Threshold sensitivity should vary tau_occ and tau_haz over a preregistered grid while keeping the alert rule fixed. Boundary tests should examine sparse point clouds, sensor noise, partial occlusion by equipment, and increasing worker density. No numerical target is asserted here; acceptable bounds must be set prospectively from site safety and compliance requirements before any empirical trial begins. [1] [3] [4]
7. Comparison with prior work
The construction studies provide the semantic and spatial basis for the proposal but stop short of a privacy model. The BIM-guided virtual-to-real segmentation framework shows that synthetic training can support point-cloud understanding without real annotations, which is relevant for privacy because it reduces the need to collect identifiable real-scene labels. The digital-twin construction monitoring study shows that segmented point clouds can populate a live site model for safety control. The present protocol extends these ideas by adding a worker-privacy constraint to the inference and exchange stages rather than treating privacy as a deployment issue external to the model. [1] [2]
The industrial federated-learning studies supply the transferable privacy machinery. One proposes hierarchical privacy budgets and poisoning-robust defenses for distributed industrial learning; another combines differential privacy, homomorphic encryption, adaptive anomaly detection, and threshold adjustment for IIoT networks; and another uses blockchain-based verification of model updates together with homomorphic encryption. The present protocol adopts the principle that privacy protection should vary with parameter sensitivity and that update exchange should be checked for integrity before aggregation. It does not assume that blockchain is necessary in construction sites, because the main privacy burden here is worker traceability through local spatial inference and model leakage rather than the absence of an auditable ledger. [3] [4] [5]
The occupant-centered building-monitoring study is a narrower but useful analog. It shows that fixed-room sensors and timetables can support personalized exposure inference without wearables, which supports the broader design principle that fixed infrastructure can replace personal devices when privacy is constrained. The construction case differs because occupancy is not scheduled and because the safety task requires spatial inference in a continuously changing environment. For that reason, the present protocol keeps the same non-wearable design principle but replaces timetable-based location assignment with digital-twin-aligned zone inference. [6]
8. Boundary conditions, failure cases, and limits of generalization
Three boundary conditions are central. First, the protocol depends on the fidelity of the digital twin or site geometry. If the mapped zones do not match current reality, the model may infer the wrong occupancy state even when the segmentation is technically correct. Second, the protocol depends on site-specific definitions of hazard. A model trained on one kind of site layout or operational rule may not generalize to another without recalibration. Third, privacy protection can weaken utility when the site uses very sparse sensing or when a hazard is visible only through features that overlap strongly with worker movement patterns. [1] [2] [3]
Failure cases are not limited to misclassification. A privacy failure occurs if an update can be used to reconstruct repeated worker paths, even when the alert itself is correct. A safety failure occurs if privacy noise suppresses the signal needed to identify a temporary exclusion-zone breach. A governance failure occurs if the system retains identifiable occupancy traces longer than required for incident review. These categories should be reviewed separately because a system can succeed in one and fail in another. This separation is important for construction monitoring, where privacy and safety are often discussed together but are not equivalent. [3] [6] [2]
Generalization from the cited evidence should be modest. The construction studies are tied to point-cloud-based site understanding and digital-twin workflows; the industrial privacy studies are tied to federated learning in sensor-rich environments. The protocol is intended to bridge those domains, not to claim that one directly validates the other. Its value lies in specifying how privacy constraints can be encoded in a construction safety pipeline and what should be measured before deployment is considered credible. [1] [2] [3] [4]
9. Conclusion
This manuscript frames worker privacy in construction safety monitoring as a design problem with explicit trade-offs rather than as a general compliance statement. The synthesis supports a fixed-infrastructure, non-wearable protocol in which point-cloud and digital-twin inference generates local safety alerts, while distributed model exchange is constrained by layer-aware privacy control, update validation, and optional encrypted aggregation for protected layers. The protocol defines variables, a single alert rule, comparators, evaluation domains, and failure cases that can be examined in future empirical studies. [1] [2] [3] [4] [5] [6]
What remains untested is whether this design can preserve operationally adequate safety monitoring, limit privacy leakage from transmitted updates and retained outputs, and remain stable across heterogeneous sites under real noise and changing geometry. That question requires prospective validation on construction data with explicit review of privacy exposure, alert correctness, and robustness. Until such validation is completed, the contribution of this manuscript is methodological: it specifies a bounded protocol for privacy-constrained worker monitoring in construction safety systems. [1] [2] [3] [4]
Contribution and availability
This manuscript proposes a research protocol for worker privacy in construction safety monitoring that combines fixed-infrastructure point-cloud inference, digital-twin-aligned zone reasoning, and privacy-preserving federated update exchange. Its contribution is a technically explicit and falsifiable method for separating safety-relevant spatial inference from unnecessary worker traceability, together with operational variables, an alert rule, a validation sequence, and boundary conditions.
Funding: Roman Vlasov’s own funds. Editorial responsibility and competing interests: Roman Vlasov is the founder and Editor-in-Chief of Smart World Journal and the creator of the BILDRY construction-planning platform. Other competing interests have not been declared.
This conceptual study reports no newly collected empirical data. Its proposed procedures require the prospective validation described above.
References
- Zou, Yiquan; Liang, Tianxiang; Nabi, Jafri Syed Riaz un; Xu, Zhendong; Zhou, Liang; Xiong, Biao (2026). A BIM-Guided Virtual-to-Real Framework for Component-Level Semantic Segmentation of Construction Site Point Clouds. Sensors. 26(1), 308. doi:10.3390/s26010308.
- Qin, Wenli (2026). Application of construction site management based on digital twin and point cloud semantic segmentation technology. PLOS One. 21(3), e0340274. doi:10.1371/journal.pone.0340274.
- Yin, Huan; Chen, Congwen; Zhang, Jingyi; Yu, Dian; Liu, Shuanggen (2026). Privacy-Preserving and Poisoning-Robust Federated Learning for Industrial IoT. Sensors. 26(16), 5297. doi:10.3390/s26165297.
- Alqazzaz, Ali (2026). SecuFL-IoT: an adaptive privacy-preserving federated learning framework for anomaly detection in smart industrial networks. Scientific Reports. 16(1), 4107. doi:10.1038/s41598-025-11883-1.
- Ahmad, Shahnawaz; Ansari, Mohd. Aquib; Mewada, Arvind; Ahmed, Gulrej (2026). Blockchain-enhanced federated learning for IoT security and privacy using the GSR-C2N model. Scientific Reports. 16(1), 26468. doi:10.1038/s41598-026-52361-6.
- Ruffa, Filippo; Iadarola, Grazia; De Capua, Alberto; De Capua, Claudio (2026). A Method for Optimized Monitoring of Indoor Air Quality in Public Buildings. Sensors. 26(14), 4559. doi:10.3390/s26144559.