PaperlayerView exact revision on arXiv

Source-linked AI summary

AquaSense-ICS: A Hybrid Real-Simulated Cyber-Physical Dataset and Leakage-Aware Benchmark for Smart Water Quality Monitoring and Attack Detection

Amal Neji, Mohamed Fourati, Lamia Chaari Fourati

arXiv:2610.32661v1cs.CR

TL;DR

Water-quality monitoring matters, but benchmark scores can be misleading when labels are derived from input features. AquaSense-ICS combines real sensing and multi-species simulation with traceable labels and leakage-aware evaluation. Under that protocol, no model beats the constant normal predictor for attack classification.

  • Problem

    Water-quality monitoring is important, yet feature-derived labels can make benchmark scores reflect recovery of the labeling rule rather than attack detection.

    Source: §1.1 · p. 1, §8.4 · p. 17

  • Method

    The paper co-registers real edge-sensor observations with EPANET-MSX multi-species simulation and evaluates attacks with leakage-aware benchmark tasks.

    Source: §1.6 · p. 2

  • Results

    No model exceeds the constant “normal” predictor for attack classification (macro-F1 0.465).

    Source: §8.7 · p. 18

  • Takeaways & Limitations

    For the legacy label, columns that are not themselves part of its definition cannot predict it, motivating tasks whose labels are not derived from inputs.

    Source: §8.4 · p. 17

    The TDS values are uncalibrated voltage-scale readings without units, and ambient air temperature is used for compensation because no immersed probe was available.

    Source: §3.3 · p. 5

Abstract

from arXiv · show

Machine-learning research on water cyber-physical systems (CPS) is constrained by datasets that are either real but chemically poor, or simulated but free of sensing noise and attacks. We present AquaSense-ICS, an open dataset and benchmark that co-registers (i) a 25.4 h recording window (8.0 h of recorded data, 26% coverage) from a Raspberry Pi 5 edge node (TDS probe and BME680 ambient sensor, 1-2 s sampling, 23,688 records), with (ii) a looped network of 13 nodes (10 junctions, 2 reservoirs, 1 tank) simulated in EPANET-MSX for six water-quality species, (iii) a parameterised library of cyber-attacks (false data injection, replay, denial of service, drift) and unintentional sensor faults with sample-level ground truth, and (iv) augmented and TimeGAN-generated sequences, flagged by an origin column. The attack, fault and water-quality labels are traceable to a seeded generator, a documented rule or a simulation input. We show that the most common way of benchmarking such hybrid data, random row splits and labels computed from the input features, yields near-perfect but uninformative scores (macro-F1 = 0.9998 for Random Forest), and we replace it with a leakage-aware protocol using chronological splits and held-out attack types rather than label reconstruction. Under this protocol, no model exceeds the constant "normal" predictor for attack classification (macro-F1 = 0.465), and the best detectors reach a mean AUROC of 0.64 on attack types unseen during training (0.35-0.93 depending on the type), leaving substantial headroom. Data and simulation files are released under CC BY 4.0 at https://doi.org/10.5281/zenodo.23390541, and attack generators and benchmark code under MIT.

1. Introduction

Water-quality monitoring and water cyber-physical security need datasets with trustworthy labels and evaluation that avoids leakage. This paper introduces a co-registered real-and-simulated dataset and benchmark, with traceable labels and leakage-aware tasks.

1.1. Motivation

  • Water is critical to health, food security, industry, and ecosystems, while water stress and unsafe drinking water underscore the importance of continuous quality monitoring.Source: §1.1 · p. 1

1.2. IoT-Based Monitoring and Cyber-Physical Systems

  • IoT enables continuous in-situ water monitoring, while cyber-physical systems integrate sensing, communication, and control but introduce security risks.Source: §1.2 · p. 1

1.3. Artificial Intelligence for Smart Water Systems

  • Machine-learning progress depends on trustworthy labels and evaluation protocols that prevent test-to-training leakage.Source: §1.3 · p. 1

1.5. Research Gap in Existing Datasets

  • No existing dataset simultaneously combines real IoT sensing, multi-species simulation, cyber-physical attacks, augmentation, interpretable labeling, and leakage-controlled evaluation.Source: §1.5 · p. 2

1.6. Proposed Dataset and Framework

  • The proposed dataset co-registers real edge sensing with EPANET-MSX multi-species simulation, while documenting that cross-source relations are synthetic by design.Source: §1.6 · p. 2, §1.8 · p. 2

1.7. Research Questions

  • The paper frames three questions about reproducible dataset construction, leakage effects and controls, and model performance on known and unseen attacks.Source: §1.7 · p. 2

1.8. Main Contributions

  • The dataset includes a seeded generator for five attack types and three fault types, with sample-level labels applied to real and simulated channels.Source: §1.8 · p. 2
  • The leakage-aware benchmark uses chronological and leave-attack-type-out splits to quantify how random splits and feature-derived labels inflate scores.Source: §1.8 · p. 2

2. Related Work

Existing smart-water datasets provide pieces of the needed evidence, but none combines real sensing, multi-species simulation, attack and fault labels, augmentation, and leakage-aware evaluation. The proposed benchmark is motivated by these gaps across real testbeds, hydraulic simulations, and attack datasets.

2.1. Overview of Existing Datasets for Smart Water Systems

  • Existing resources span real IoT sensing, hydraulic simulation, cyber-physical attacks, and hybrid datasets, categories compared alongside the proposed benchmark.Source: §2.1 · p. 2

2.2. Real Testbed and Sensing Datasets

  • SWaT and WADI provide real sensor measurements and labeled attacks, but remain controlled laboratory testbeds with few water-quality analyzers and without network-scale chemical variables.Source: §2.2 · p. 2
  • WaterBench offers deployed IoT measurements but lacks attack scenarios, quality labels, and augmentation, while reviews identify a broader shortage of unified open resources.Source: §2.2 · p. 3

2.3. Hydraulic Simulation Benchmarks

  • EPANET-MSX supports multi-species reaction modeling, but its datasets remain scarce and are not integrated with real observations or cybersecurity scenarios.Source: §2.3 · p. 3
  • BattLeDIM and LeakDB provide simulation-based leakage benchmarks without water-quality modeling, while BattLeDIM also lacks cyber-attacks and real sensing integration.Source: §2.3 · pp. 3–4

2.4. Cyber-Physical Attack Datasets

  • BATADAL provides labeled simulated attack events, but lacks real IoT measurements, water-quality parameters, and time-series augmentation; KIOS datasets also lack security enrichment and multi-species chemistry.Source: §2.4 · p. 4

2.5. Limitations of Existing Resources and Motivation

  • No existing public dataset combines real edge sensing, multi-species hydraulic and chemical simulation, parameterized attack and fault labels, separate augmentation, interpretable quality labels, reproducibility, and leakage-aware evaluation.Source: §2.5 · p. 4

3. Measurement Architecture and Real Data Acquisition

The platform combines Raspberry Pi edge acquisition with water-quality and ambient sensing, but its released TDS value is an uncalibrated, unitless scaled voltage rather than a calibrated concentration.

3.1. Proposed IoT Architecture

  • A Raspberry Pi 5 handles sensor interfacing, acquisition, local storage, and wireless transmission in the monitoring platform.Measurements are stored locally as CSV and can be accessed remotely over Wi-Fi.Source: §3.1 · p. 4

3.2. Sensors and Hardware Configuration

  • The DAR-02-W40 TDS sensor measures total dissolved solids in water, reflecting overall salinity and impurity concentration, and outputs an analog voltage proportional to ion concentration, digitized by the ADS1115 ADC before transmission to the Raspberry Pi.Source: §3.2 · p. 4

3.3. TDS Measurement and Limitations

  • The released tds_value is an uncalibrated, unitless scaled voltage, not a calibrated TDS concentration.The probe was not calibrated against standard solutions, raw voltage was not retained, and ambient air temperature is used for compensation because no immersed temperature probe was available.Source: §3.3 · pp. 4–5

3.4. Measurement Ranges and Operating Thresholds

The study defines operating thresholds for sensor-based anomaly labels, collects a short interrupted recording with controlled ambient perturbations, and provides a rerunnable preprocessing pipeline.

3.4. Measurement Ranges and Operating Thresholds

  • TDS below 600 ppm is treated as acceptable, with excellent quality below 300 ppm and good quality from 300–600 ppm; WHO sets no health-based TDS guideline value.The study also sets temperature, humidity, and gas-resistance operating ranges, and uses the thresholds to generate an ambient-anomaly flag distinct from the water-quality label.Source: §3.4 · p. 5

3.5. Software Environment

  • The acquisition stack uses Python and sensor libraries, stores timestamped CSV records locally, and supports remote Raspberry Pi monitoring.The repository pins exact package versions in requirements.txt.Source: §3.5 · pp. 5–6

3.6. Data Acquisition Methodology

  • The 25.4-hour collection window contains 8.0 hours of samples (26% coverage), interrupted by 23 gaps longer than 60 seconds.Within recorded periods, samples are most often 1–2 seconds apart; the campaign is described as an initial proof-of-concept.Source: §3.6 · p. 6

3.7. Experimental Perturbation Scenarios

  • Smoke, nearby water, and a candle were used to perturb gas resistance, humidity, and temperature, respectively, but lack of event times prevents their use as ground-truth event labels.These disturbances affect the ambient environment rather than the water itself.Source: §3.7 · p. 6

3.8. Data Preprocessing Pipeline

  • The rerunnable preprocessing pipeline retains observed channels and quality flags, applies no imputation, and uses 30-sample windows with stride 5 labelled by the final sample.Normalization is fitted on the training split only; gas resistance is log-transformed before z-scoring.Source: §3.8 · p. 6

4. Hydraulic and Chemical Simulation using EPANET-MSX

EPANET-MSX supplies multi-species chemical dynamics for a 13-node distribution zone, with coupled reaction rules and released simulation inputs producing hourly states for co-registration with real IoT data.

4.1. EPANET-MSX Simulation Framework

  • EPANET-MSX extends hydraulic simulation to coupled multi-species reactions in bulk water and pipe-wall domains.The modeled processes include chlorine decay, pH variation, turbidity evolution, dissolved oxygen consumption, conductivity variation, THM formation, and wall chlorine demand.Source: §4.1 · pp. 6–7

4.2. Hydraulic Network Topology

  • The simulated single pressure zone has 13 nodes and 16.8 km of pipes, combining looped and branched paths with sources, a tank, a pump, and control valves.The size supports tracing anomalies to physical propagation paths, and the network and reaction files are released for application to larger systems.Source: §4.2 · p. 7

4.3. Reaction Kinetics and Multi-Species Interaction Model

  • Chlorine depletion couples bulk degradation, dissolved-oxygen consumption, and pipe-wall biofilm demand, while THM forms as a chlorine-dependent secondary product.The model treats pH as relaxing toward 7.4, turbidity as settling without resuspension, and conductivity as first-order decay rather than a conservative tracer.Source: §4.3 · p. 8

4.4. Simulation Parameters and Output Dataset

  • The simulation exports 24 hourly states of six species at 13 nodes, including two-hour chlorine and turbidity pulse injections and localized initial perturbations.The complete synthetic time series supports spatial analysis of contaminant propagation and temporal anomaly detection, and is co-registered with real IoT measurements.Source: §4.4 · p. 8

5. Data Augmentation and Cyber-Physical Perturbations

A seeded perturbation pipeline supplies training-only augmentation and synthetic sequences alongside cyberattacks and sensor faults with sample-level labels. These additions target data scarcity and realistic cyber-physical variation while keeping generated data out of validation and test splits.

5.1. Time-Series Augmentation

  • Five time-series transformations add noise, gain changes, temporal warping, partial-window gaps, or reordered segments to four sensor channels.Jittering uses σ = 0.05; scaling uses [0.9, 1.1]; window slicing retains 90% of a series.Source: §5.1 · p. 8
  • Augmentation and TimeGAN or bootstrap sequences are released separately and applied to training data only, avoiding row-aligned copies that can leak labels or mismatch timestamps.The perturbations are implemented in one seeded generator, aquasense.perturb.Source: §5 · p. 8, §5.1 · p. 9

5.2.1. Bootstrapping

  • Bootstrap generation doubles the data by iid resampling, but does not preserve temporal order or short-range autocorrelation.It also adds Gaussian noise to TDS and a linear drift from 0 to 1.Source: §5.2.1 · p. 9
  • 0.037 vs. 0.172: the simple bootstrap has a lower discriminative score than TimeGAN on held-out real data.The score is the absolute distance from 0.5 accuracy of a classifier separating real from synthetic; lower is better.Source: §5.2.1 · p. 10

5.3. Cyber-Physical Attack Modeling

  • The attack library models FDI, replay, DoS, and progressive drift, with mechanisms ranging from altered values to stale or missing measurements.FDI leaves the physical system unchanged; replay reuses a past window, and drift adds growing error.Source: §5.3 · p. 10
  • Fourteen attack segments affect 10.0% of real samples; several attacks stay within clean channel ranges, while DoS uses frozen values or missing readings rather than zeros.The released streams retain sample-level attack type, ID, and channel labels.Source: §5.3 · pp. 10–11

5.4. Realistic Noise and Anomalies

  • Three separately labeled sensor-fault types model white-noise fluctuations, isolated spikes, and full-scale extremes, from small variability to complete faults.Noise affects all sensor channels, while spikes and extremes are injected one channel at a time.Source: §5.4 · p. 11

5.5. Section Summary

  • The section combines statistical and temporal augmentation, synthetic generation, attack modeling, and sensor-noise injection as a four-level robustness pipeline.Generated and augmented data are supplied in separate files.Source: §5 · p. 8, §5.5 · p. 11

6. Temporal Fusion and Water Quality Labeling

The paper aligns real sensor records with hourly EPANET-MSX states and combines them with separately identified synthetic sequences. Water-quality labels are rule-based, but the resulting class is constant among samples with a matched simulation state, so it is not used as a benchmark target.

6.2.1. Join Selection

  • The fusion preserves one row per real sample and keeps observed chlorine and turbidity channels distinct from their hourly EPANET counterparts.The fused table contains 23,688 rows, with simulated columns renamed using the suffix _epanet where names collide.Source: §6.2.1 · p. 12

6.2.2. Temporal Propagation

  • The backward as-of join matches 20,795 of 23,688 real samples to hourly simulation states; 2,893 remain missing, including samples before the first and after the last state.The join uses a tolerance of one reporting interval, 3,600 s.Source: §6.2.1 · p. 11, §6.2.2 · p. 12
  • The fused table contains the J3 chlorine pulse at τ = 7 h, matched by 1,427 samples, but the J7 turbidity pulse at τ = 15 and 16 h matches no real sample.The J7 pulse remains visible only in simulated_data.csv; the approach assumes quasistationary network behavior over one hour and neither invents values nor moves them across nodes.Source: §6.2.2 · p. 12

6.2.3. Synthetic Sequences

  • TimeGAN contributes 23,688 training-only rows with four sensor channels but no simulated state, water-quality class, or attack and fault labels.The resulting file has 47,376 rows; synthetic rows are marked origin = timegan.Source: §6.2.3 · p. 12, §6.3.4 · p. 14

6.2.4. Post-Fusion Cleaning •

  • The released real rows contain five attack classes covering 10.0% of samples and three fault classes covering 10.7%.The most frequent attack is replay (1,044 samples); white noise is the most frequent fault (2,368 samples).Source: §6.2.4 · p. 13

6.3. Automatic Water Quality Label Generation

  • The TDS readings cannot be verified as ppm, so the ppm thresholds cannot meaningfully discriminate samples; every recorded value receives the maximum TDS score.The raw probe voltage needed to verify conversion was not retained.Source: §6.3 · p. 13

6.3.1. Composite Scoring Approach

  • The water-quality score combines six physicochemical indicators at J3, with each indicator evaluated against optimal and acceptable thresholds.The score ranges from 0 to 12, although the dataset maximum is 11.Source: §6.3.2 · p. 13, §6.3.1 · p. 13

6.3.3. Class Assignment with Fixed Thresholds

  • 20,795 real samples with simulated states all receive the “Very Good” class, while 2,893 unmatched samples receive no class; wq_class is therefore excluded as a benchmark target.The matched samples all score 11, partly because TDS always receives its maximum score and chlorine never reaches its optimal range.Source: §6.3.3 · p. 13

7. Exploratory Data Analysis

Exploratory analysis finds associations among simulated species but only weak dependence among sensor channels. It also quantifies how injected attacks and faults alter the real sensor stream.

7.1. Correlation Analysis

  • The strongest sensor-channel correlation is only ρ = +0.29, and no sensor pair reaches the |ρ| > 0.40 reporting threshold.Real–simulated correlations are not computed because the channels are co-registered, not coupled.Source: §7.1 · p. 15, Figure 6 · p. 16
  • All six reported correlations above |ρ| = 0.40 are among simulated species, and the strongest involve dissolved oxygen, conductivity, and THM.These correlations describe spatial structure across 13 nodes, not independent observations; the chlorine–THM association is −0.44 with a 95% interval including zero.Source: §7.1 · pp. 14, 16

7.2. Effect of Attacks and Faults on the Sensor Channels

  • At least one sensor channel differs from the raw record in 3,568 real samples (15.1%), with white noise affecting all four channels and other perturbations affecting selected channels.Rows without an attack or fault are identical to the raw record.Source: §7.2 · p. 16

8. Machine Learning Model Benchmark

The leakage-aware evaluation replaces label reconstruction with chronological attack classification and leave-one-attack-type-out detection. No model beats the constant-normal classification baseline, and unseen-attack detection varies substantially by attack type.

8.4. Analysis of Results

  • The first-release near-perfect scores reflect label leakage, near-duplicate inputs, temporal neighbors across random splits, and an attack-label feature—not evidence of generalization.A tree using the eight label-defining columns alone recovers the legacy label with 0.9947 accuracy.Source: §8.4 · p. 17, Table 23 · p. 17
  • On the current release, Random Forest macro-F1 falls from 0.976 with a random split to 0.341 with a chronological split.Removing augmented copies alone barely changes the score because the label-defining columns remain.Source: §8.4 · p. 17, Table 24 · p. 18

8.6. Leakage-Aware Benchmark: Tasks

  • The leakage-aware protocol uses chronological splits and leave-one-attack-type-out detection, with held-out attack windows scored against normal windows from validation and test periods.The authors note that differing periods between positive and negative windows may contribute temporal shift to the AUROC.Source: §8.6 · p. 18

8.7. Leakage-Aware Benchmark: Results

  • Deep models, including 1D-CNN, LSTM, Transformer, and LSTM autoencoder, were not evaluated, and no on-device measurements were performed.The candidate-model comparison is qualitative for size and interpretability.Source: §8.6 · p. 18, §8.7 · p. 18, Table 28 · p. 19
  • 0.465 macro-F1: no model exceeds the constant “normal” predictor; mean unseen-attack AUROC is 0.64 for Random Forest and Isolation Forest.Detection is above chance for DoS, FDI, and drift with these models, but at or below chance for FDI-stealth and replay; the FDI estimate uses a single segment.Source: §8.7 · p. 18, Table 26 · p. 19
  • Random Forest and XGBoost label every test window normal, while logistic regression scores 0.417 macro-F1 with a 19.3% false-alarm rate.Isolation Forest and one-class SVM raise false alarms but detect no replay segment.Source: §8.7 · p. 18, Table 26 · p. 19
  • Training-window augmentation and TimeGAN leave Random Forest at 0.465 macro-F1, and TimeGAN-only training cannot be evaluated because its rows have no attack labels.The forest's predictions remain constant across configurations.Source: §8.7 · p. 18, Table 27 · p. 18

8.8. Final Model Selection

  • Random Forest is selected for XAI analysis because it ties for best macro-F1 and has the highest mean unseen-attack AUROC among evaluated supervised models.Its cited values are 0.465 macro-F1 and 0.64 mean AUROC.Source: §8.8 · p. 19

9. SHAP Methodology (SHapley Additive exPlanations)

The section uses SHAP to examine a legacy-label classifier and finds its apparent importance is dominated by label-defining inputs and their augmented copies. Grouping correlated features clarifies their shared importance, while the analysis motivates label-independent inputs and leakage-aware evaluation.

9. SHAP Methodology (SHapley Additive exPlanations)

  • Under the v1 protocol, the Random Forest (200 trees, random_state=42), retrained on 28 inputs, reached 0.9976 accuracy and 0.9967 macro-F1 on the test split.The 23,688 synthetic rows lacked values for 24 of the 28 inputs; these were imputed with the training-set median before standardisation.Source: §9 · p. 20

9.2. Analysis of Engineered Feature Redundancy

  • After removing the 20 copies, the eight label-defining inputs still yield 0.9904 accuracy and 0.9866 macro-F1, so removing duplicated features does not remove label leakage.On real rows only, macro-F1 is 0.9998; individual column rankings vary, while the group-level result is more stable.Source: §9 · p. 20

9.1. Global Feature Importance Analysis (v1 Protocol)

  • Temperature is the dominant predictor, with a SHAP value 36% higher than tds_value; all four raw sensor variables among the top nine define the legacy label.None of the node-level simulated variables appears among the nine most influential features, consistent with label reconstruction rather than detection of a physical water-quality signal.Source: §9.1 · p. 20

9.2. Analysis of Engineered Feature Redundancy

  • Six of the nine top-ranked variables are augmented copies, and correlated copies divide importance, so source-column groups—not individual ranks—are the relevant comparison.By grouped mean |ϕ|, temperature accounts for 50.1% and tds_value for 33.0%; the four Cl/TURB columns together account for 0.3%.Source: §9.2 · p. 20, §9.1 · p. 20

9.3. Design Implications and Practical Insights

  • The analysis recommends auditing datasets with explainability and deriving labels without sharing input columns, while compact-feature detection on leakage-aware tasks remains untested.SHAP shares credit among correlated features, which must be considered when interpreting rankings.Source: §9.3 · p. 20

10. Data Availability and Reuse

The release provides the data, models, documentation, and benchmark materials for reuse, with distinct licenses for the dataset and code.

10. Data Availability and Reuse

  • The data tables are archived on Zenodo under CC BY 4.0, while code, EPANET/EPANET-MSX files, documentation, and official splits are available through the GitHub repository.The files are provided as CSV; data are CC BY 4.0 and code is MIT.Source: §10 · pp. 20–21
  • The dataset is intended for anomaly and attack detection, forecasting, synthetic-data, and domain-adaptation research, but not for potability certification or physical conclusions from real–simulated channel relations.Future releases are planned under the same concept DOI and schema to preserve comparability.Source: §10 · p. 21

11. Conclusion and Future Directions

AquaSense-ICS combines real edge sensing, EPANET-MSX simulation, and traceable attack and fault labels, then evaluates detection with leakage-aware splits. Results show that random feature-derived-label scores are inflated, while attack detection remains difficult under the leakage-aware protocol; the authors identify data and evaluation boundaries and propose expansions.

11.1. Summary of Contributions

  • Under chronological and leave-attack-type-out evaluation, no model exceeds the constant predictor in attack-classification macro-F1 (0.465), and the best detectors reach mean AUROC 0.64 on unseen attack types.Detection is above chance for DoS, FDI, and drift, and at or below chance for replay and stealthy FDI.Source: §11.1 · p. 21
  • The open dataset co-registers a real 25.4 h recording, with 8.0 h recorded, and a six-species EPANET-MSX model, adding seeded attack and fault generators with sample-level labels.Attack and fault labels trace to a seeded generator, while water-quality labels trace to documented thresholds; that water-quality label is constant on available data.Source: §11.1 · p. 21
  • Random splits with feature-derived labels inflate macro-F1 to 0.9998; on the current release, the same classifier drops from 0.976 to 0.341 macro-F1 under a chronological split.The benchmark is presented as a template for evaluating other hybrid real–simulated datasets.Source: §11.1 · p. 21

11.2. Limitations

  • The real acquisition is a short, single-indoor-site deployment, and its limited geographic coverage may not capture large-network spatial heterogeneity or seasonal effects.Only 8.0 hours were recorded within the 25.4-hour window.Source: §11.2 · p. 21
  • The real and simulated systems share a clock but are independent, attacks affect recorded data rather than closed-loop control consequences, and the simulated topology is not a specific real infrastructure.These boundaries limit physical coupling and direct site-specific engineering applicability.Source: §11.2 · p. 21
  • Water-quality labels are rule-based and lack laboratory grab-sample validation, while the label is constant on available data and cannot serve as a classification target.Only TDS is measured in water; temperature, humidity, and gas resistance describe ambient air, and no immersed temperature probe was available for TDS compensation.Source: §11.2 · p. 21
  • Evaluation of unseen attack types is indicative because the release has 14 segments, with only one segment each for FDI and stealthy FDI.Attack types therefore cannot be distributed across training, validation, and test splits.Source: §11.2 · p. 22

11.3. Future Research Directions

  • Future work proposes multiple field sites and additional sensors, larger networks, expanded attack scenarios, seasonal coverage, improved labeling, and multi-site federated or foundation-model studies.The present single-site version cannot meaningfully test federated learning across sites.Source: §11.3 · p. 22

CRediT authorship contribution statement

The authors’ contributions span conceptualization, data curation, validation, software, methodology, investigation, writing, and supervision.

CRediT authorship contribution statement

  • A.N. handled conceptualization, methodology, software, investigation, and the original draft; M.F. and L.C.F. contributed data curation and validation, with L.C.F. also supervising.M.F. and L.C.F. contributed to review and editing.Source: CRediT authorship contribution statement · p. 22
Loading 2610.32661v1…