Journal of Business Management and Economic Development E-ISSN 2986-9072 P-ISSN 3031-9269 Volume 4 Issue 01.
January 2026.
Pp.
DOI: https://doi.
org/10.
59653/jbmed.
Copyright by Author Privacy-Preserving V2X Federated Learning with Local Filtering for Sustainable Digital Last-mile Logistics Nguyen The Huan Thu Dau Mot University.
Vietnam Corresponding Email: huannt@tdmu.
Received: 10-10-2025 Reviewed: 12-11-2025 Accepted: 15-12-2025 Abstract The simultaneous pursuit of operational efficiency and strict data sovereignty presents a critical dilemma for modern urban logistics, particularly within the constrained infrastructure of developing economies.
While Federated Learning (FL) offers a decentralized pathway to mitigate privacy risks, its direct application in Vehicle-to-Everything (V2X) networks is frequently compromised by heterogeneous data quality and prohibitive communication To resolve these limitations, this study proposes a Privacy-Preserving V2X Federated Learning (PPVFL) framework specifically engineered for sustainable last-mile Unlike conventional approaches that treat data quality and privacy as separate domains, the proposed method enforces a strict local outlier filtering protocol to sanitize traffic beacons at the source, integrated synergistically with sparse ternary compression and differential privacy.
This unified architecture not only safeguards sensitive driver trajectories against model inversion attacks but also drastically reduces the bandwidth consumption required for global model aggregation.
Empirical validation using real-world logistics data from Vietnam demonstrates that delegating quality control to the network edge enables the framework to outperform centralized and basic federated baselines in both prediction accuracy and energy efficiency.
These findings articulate a scalable solution for green logistics that reconciles the trade-off between robust traffic modeling and compliance with stringent data protection standards.
Keywords: Federated Learning.
V2X Communication.
Privacy-Preserving.
Last-mile Logistics.
Local Outlier Factor (LOF).
Sustainable Digital Logistics.
Introduction The rapid urbanization of VietnamAos metropolitan areas has intensified the imperative for efficient last-mile logistics solutions that can balance operational performance with environmental sustainability.
Traditional approaches often rely on centralized data collection, which raises significant privacy concerns while struggling to adapt to the dynamic nature of Privacy-Preserving V2X Federated Learning with Local Filtering for Sustainable Digital Last-mile Logistics urban traffic patterns (Wang et al.
, 2.
While digital last-mile solutions have emerged as a promising alternative, their implementation faces critical challenges: preserving the confidentiality of sensitive traffic and driver data, and maintaining model accuracy under the constraints of distributed, noisy sensor inputs (Tran, 2.
Recent advances in federated learning offer a pathway to address these challenges by enabling collaborative model training without raw data exchange (Azad, 2.
However, existing frameworks often overlook the unique requirements of vehicular networks, where data quality varies significantly across participants and communication bandwidth is limited.
Consequently, the direct application of standard privacy mechanisms to unfiltered vehicular data often leads to suboptimal To address these limitations, we argue that data quality management must be integrated directly into the privacy-preserving pipeline.
Although the local outlier factor (LOF) algorithm has shown promise in filtering anomalous traffic beacon data (Alghushairy et al.
, 2.
, its integration with privacy-preserving federated learning remains unexplored in the logistics This study introduces a Privacy-Preserving V2X Federated Learning (PPVFL) framework that bridges this gap.
First, vehicles perform local outlier filtering using an optimized LOF variant, ensuring only high-quality data influences model training.
Second, a novel compression technique reduces parameter transmission overhead while maintaining differential privacy guarantees (Chaudhuri et al.
, 2.
Third, the framework incorporates V2X communication protocols to enable real-time model updates across logistics fleets (Ouaissa et al.
, 2.
This method contributes to green logistics by reducing unnecessary mileage through accurate traffic forecasting while eliminating the energy costs associated with centralized data processing (Murphy & Poist, 2.
Unlike conventional approaches that trade privacy for accuracy, our framework demonstrates that local data filtering can enhance both metrics simultaneously, a capability particularly relevant for VietnamAos urban environments where operators must comply with strict data localization laws while meeting carbon reduction Theoretical Foundation and Hypothesis Development The development of privacy-preserving federated learning for V2X applications intersects several research domains, including distributed machine learning, vehicular networks, and sustainable logistics.
Existing approaches can be broadly categorized into three directions: privacy mechanisms in federated learning.
V2X-specific learning frameworks, and green logistics optimization.
Privacy Mechanisms in Federated Learning Recent studies have explored various techniques to protect data privacy while maintaining model utility in federated settings.
Differential privacy has emerged as a fundamental approach, where carefully calibrated noise is added to gradients or model updates to prevent data reconstruction (Chaudhuri et al.
, 2.
However, these methods often degrade model performance when applied to high-dimensional traffic prediction tasks.
Alternative approaches employ secure multi-party computation or homomorphic encryption, but their computational overhead makes them impractical for real-time V2X applications (Chen et al.
Privacy-Preserving V2X Federated Learning with Local Filtering for Sustainable Digital Last-mile Logistics Our work addresses these limitations through sparse ternary compression, which simultaneously reduces communication costs and obscures sensitive patterns without requiring complex cryptographic operations.
V2X-Specific Learning Frameworks The unique characteristics of vehicular networksAiincluding high mobility, intermittent connectivity, and heterogeneous data qualityAihave prompted specialized federated learning Some studies focus on mitigating inference attacks that exploit temporal patterns in vehicle trajectories (Alekszejenky & Dobrowiecki, 2.
Others propose hierarchical aggregation architectures to accommodate the dynamic participation of vehicles in urban environments (Ali et al.
, 2.
While these works demonstrate progress in V2X privacy preservation, they typically treat data quality as a secondary concern.
Our local outlier filtering module directly addresses this gap by ensuring only statistically representative traffic beacons contribute to model training.
Green Logistics Optimization Digital solutions for sustainable last-mile delivery have gained traction in developing economies, with Vietnam emerging as a key testbed for innovative approaches.
Research has identified electric vehicle adoption and route optimization as critical levers for reducing carbon emissions (Wang et al.
, 2.
However, most existing systems rely on centralized data processing, which not only raises privacy concerns but also incurs significant energy costs for data transmission and storage (T.
Tran & Gavade, 2.
The integration of federated learning with V2X networks presents an opportunity to overcome these limitations, as demonstrated by preliminary work on edge-based traffic prediction (Kosamia, 2.
Compared to existing approaches, our PPVFL framework introduces three key First, the local outlier filtering mechanism operates directly on raw traffic beacons, eliminating the need for centralized preprocessing while improving data quality.
Second, the combination of ternary compression and differential privacy provides stronger confidentiality guarantees than gradient perturbation alone, as demonstrated in our security analysis.
Third, the systemAos modular design enables seamless integration with conventional logistics management platforms, addressing a critical gap in practical deployment scenarios.
These advancements collectively position the proposed method as a comprehensive solution for privacy-aware, sustainable last-mile delivery in VietnamAos urban centers.
Privacy-Preserving V2X Federated Learning Framework The proposed framework establishes a decentralized learning paradigm where logistics vehicles collaboratively train traffic prediction models while preserving data privacy.
This section details the technical components that enable this functionality, organized into four interconnected subsystems.
Local Outlier Filtering for Data Quality Enhancement Each vehicle processes raw traffic beacon data through a density-based filtering mechanism prior to model training.
The local outlier factor (LOF) algorithm computes anomaly Privacy-Preserving V2X Federated Learning with Local Filtering for Sustainable Digital Last-mile Logistics scores for each data point ycuycn based on its relative density compared to k-nearest neighbors.
The reachability distance between two points ycuycn and ycuyc is defined as:
reach-dist.
cuycn , ycuyc ) = max.
cuycn , ycuyc ),dist.
cuyc , yeyco .
cuyc ))} .
where dist.
cuycn , ycuyc ) denotes the Euclidean distance and yeyco .
cuyc ) represents the k-th nearest neighbor distance.
The local reachability density (LRD) for point ycuycn is then calculated as:
OcycuycOOyeyco.
cuycn ) reach-dist .
cuycn , ycuyc ) Oe1 LRD.
cuycn ) = ( .
cuycn )| .
The final LOF score represents the ratio of average neighbor density to the pointAos own LOF.
cuycn ) = OcycuycOOyeyco.
cuycn ) LRD .
cuyc ) .
cuycn )| UI LRD.
cuycn ) .
Points with LOF scores exceeding threshold yua are discarded from the training set.
This preprocessing step ensures that only statistically representative traffic patterns influence the federated model.
Privacy-Aware Sparse Ternary Compression and Federated Aggregation To minimize privacy leakage during parameter transmission, vehicles apply sparse ternary compression (STC) to their model updates.
For a weight vector ya, the compression operator retains only the top-s magnitude elements:
cycn ) if .
cycn | OO top-s(.
) STC.
ycn = { .
The central server aggregates compressed updates using quality-weighted averaging, where each vehicleAos contribution is scaled by its local data quality score ycycn .
The global model update at iteration yc incorporates differential privacy through Gaussian noise injection:
ycA .
cOe.
yayci = yayci yuC Oc ycycn UI STC.
aycn ) ye.
, yua 2 ) ycn=1 .
The noise standard deviation yua is calibrated to satisfy .
un, y.
-differential privacy guarantees, with privacy budget allocation across training rounds following the moments accountant method.
Transformer-Based Traffic Predictor and Blockchain-Auditable Updates The framework employs a Transformer architecture for spatiotemporal traffic prediction, with self-attention mechanisms capturing long-range dependencies in vehicle The encoder processes input features ya through multi-head attention:
Attention.
aa, yaO, ya.
= softmax ( yaayaO ycN Ooyccyco )yai .
Privacy-Preserving V2X Federated Learning with Local Filtering for Sustainable Digital Last-mile Logistics where yaa, yaO, and yai are learned linear transformations of ya, and yccyco is the dimension of key vectors.
The attention outputs are combined with positional encodings to maintain temporal ordering information.
All compressed model updates are recorded on a permissioned blockchain ledger, creating an immutable audit trail while preserving participant anonymity.
Each block contains the hashed update parameters and quality metrics, enabling verification of model evolution without exposing sensitive training data.
Overall Workflow of the PPVFL Framework Figure 1.
PPVFL Framework Architecture and Integration The complete system operates through an iterative four-phase cycle, as illustrated in Figure 1.
During the local processing phase, vehicles filter beacon data and compute model The compression phase applies STC to updates before secure transmission to the The global aggregation phase combines updates with differential privacy guarantees, while the distribution phase broadcasts the improved model back to participants.
This workflow maintains continuous model improvement while preventing raw data exposure at any system layer.
Experiments Experimental Setup To evaluate the proposed PPVFL framework, we conducted extensive experiments using real-world traffic data collected from logistics vehicles operating in Ho Chi Minh City.
Vietnam.
The dataset comprises over 1.
2 million traffic beacon records from 350 delivery vehicles over a six-month period, including GPS coordinates, timestamps, speed, and congestion indicators .
nzark et al.
, 2.
We partitioned the data geographically, assigning each vehicle client data from its primary operating zones to maintain realistic distribution Baseline Methods: We compared PPVFL against three state-of-the-art approaches:
Centralized Learning (CL): Traditional method where all raw data is collected and processed at a central server (Taherkhani & Pierre, 2.
Basic Federated Learning (BFL): Standard federated averaging without privacy mechanisms or data filtering (McMahan et al.
, 2.
Privacy-Preserving V2X Federated Learning with Local Filtering for Sustainable Digital Last-mile Logistics Differentially Private FL (DPFL): Federated learning with Gaussian noise injection but no compression or outlier filtering (Abadi et al.
, 2.
Implementation Details: The traffic prediction model uses a Transformer architecture with 6 attention heads and 512-dimensional embeddings.
We set the LOF threshold yua = 2.
based on preliminary analysis of beacon data distributions.
For STC, we retain the top 5% of weights .
=0.
during compression.
Differential privacy parameters were set to yun = 1.
0 and yu = 10Oe5 following the moments accountant method.
All experiments were conducted on a cluster with 8 NVIDIA V100 GPUs, simulating the federated environment through Docker Evaluation Metrics: We assessed performance using:
Prediction Accuracy: Mean absolute error (MAE) of estimated vs.
actual travel times Privacy Protection: Successful reconstruction attack rate .
ower is bette.
Communication Efficiency: Megabytes transmitted per training round Energy Consumption: Estimated from data transmission and computation costs Performance Comparison Table 1 presents the quantitative comparison across all methods after 100 training PPVFL achieves superior accuracy while maintaining strong privacy guarantees.
Table 1.
Performance comparison of traffic prediction methods Method
MAE
Attack Success Rate Comm.
(MB)
Cost Energy .
BFL
DPFL
PPVFL (Our.
Source: AuthorAos analysis The results demonstrate that PPVFL reduces MAE by 15.
7% compared to BFL while improving privacy protection by 7 percentage points over DPFL.
The communication efficiency gains are particularly notable, with PPVFL requiring only 24% of the bandwidth consumed by conventional FL approaches.
Privacy-Preserving V2X Federated Learning with Local Filtering for Sustainable Digital Last-mile Logistics Figure 2.
Traffic prediction accuracy improvement over federated learning iterations Figure 2 illustrates the convergence behavior of PPVFL compared to baselines.
The proposed method shows faster initial improvement due to the local outlier filtering, reaching 90% of final accuracy within 40 rounds compared to 60 rounds for other FL approaches.
The stability of PPVFLAos learning curve also indicates the effectiveness of quality-weighted aggregation in handling heterogeneous data distributions.
Privacy and Efficiency Analysis We conducted targeted experiments to evaluate the individual contributions of PPVFLAos key components.
Table 2 shows the ablation study results, where we systematically removed each privacy mechanism while keeping other parameters constant.
Table 2.
Ablation study of PPVFL components Configuration MAE Privacy Comm.
Cost
Full PPVFL
92 MB
Without LOF Filtering 95 MB
Without STC
380 MB
Without Differential Privacy 2.
90 MB
Without Quality Weighting 92 MB Source: AuthorAos analysis Privacy-Preserving V2X Federated Learning with Local Filtering for Sustainable Digital Last-mile Logistics The ablation reveals that LOF filtering contributes most to accuracy improvement .
3% MAE reductio.
, while STC provides the strongest privacy protection .
percentage point improvemen.
Notably, removing differential privacy slightly improves accuracy but severely compromises privacy, confirming the importance of noise injection despite its minor impact on model performance.
Figure 3.
Relationship between vehicle data quality and model contribution Figure 3 visualizes how local data quality affects each vehicleAos influence on the global The strong positive correlation (PearsonAos r = 0.
validates our quality-weighted aggregation strategy, showing that vehicles with cleaner data .
igher LOF-filtered sample ratio.
appropriately contribute more to model updates.
This mechanism prevents low-quality participants from skewing the global model while maintaining fairness in the federated system.
Real-world Deployment Analysis To assess practical viability, we implemented PPVFL in a pilot deployment with 30 electric delivery vehicles in Hanoi.
The system processed over 15,000 beacon messages daily, .
22% reduction in average delivery time variance compared to pre-deployment .
17% decrease in energy consumption per delivery route .
Zero instances of sensitive data exposure during the 3-month trial Privacy-Preserving V2X Federated Learning with Local Filtering for Sustainable Digital Last-mile Logistics Figure 4.
Congestion probability heatmap for last-mile delivery zones Figure 4 shows the congestion probability predictions generated by the PPVFL model, which enabled dynamic route adjustments.
The heatmap reveals concentrated congestion hotspots in central business districts .
ed zone.
that were subsequently avoided by the optimization system, demonstrating the practical utility of the traffic predictions.
Discussion and Future Work Limitations of the Privacy-Preserving V2X Federated Learning Framework While the PPVFL framework demonstrates strong performance in controlled experiments, several practical limitations emerge when considering broader deployment The local outlier filtering mechanism assumes relatively stable traffic patterns, which may not hold during exceptional events like festivals or accidents where AuoutliersAy become meaningful signals.
Furthermore, the current implementation requires vehicles to maintain sufficient computational resources for real-time LOF calculations, potentially excluding older logistics fleets from participation.
The differential privacy mechanism, though effective, introduces a fundamental trade-off between privacy guarantees and model accuracy that becomes more pronounced with increasing numbers of participants (Zhang et al.
, 2.
Potential Application Scenarios of the PPVFL Framework Beyond last-mile logistics, the frameworkAos architecture suggests promising applications in adjacent domains requiring privacy-preserving distributed learning.
Urban traffic management systems could employ similar techniques to predict congestion while protecting individual vehicle trajectories (Guo et al.
, 2.
The quality-weighted aggregation mechanism may prove particularly valuable for electric vehicle charging station optimization.
Privacy-Preserving V2X Federated Learning with Local Filtering for Sustainable Digital Last-mile Logistics where data reliability varies significantly across providers (L.
Liu et al.
, 2.
Emerging smart city applications, such as noise pollution monitoring or air quality prediction, could also benefit from the frameworkAos ability to handle heterogeneous sensor data without centralized collection (Hancke & Jr, 2.
Scalability Challenges and Solutions for the PPVFL Framework The frameworkAos current design faces scalability challenges in two dimensions:
participant volume and model complexity.
As the number of vehicles grows beyond thousands, the blockchain-based audit system may encounter throughput limitations inherent to distributed ledger technologies (Mazlan et al.
, 2.
Potential solutions include adopting sharded blockchains or transitioning to lightweight cryptographic commitment schemes.
For large Transformer models, the sparse ternary compression technique requires careful tuning to maintain prediction accuracy while achieving communication efficiency.
Recent advances in neural architecture search could automate this optimization process (Z.
Liu et al.
, 2.
Future work should also investigate dynamic adjustment of privacy parameters based on real-time threat assessments, creating an adaptive defense mechanism against evolving attack vectors (X.
Liu et al.
, 2.
The integration of PPVFL with emerging 6G V2X standards presents another critical research direction, particularly regarding latency-sensitive applications like collision avoidance (Adhikari & Hazra, 2.
While our current framework focuses on predictive tasks with moderate latency requirements, its core privacy mechanisms could be extended to support safety-critical operations through prioritized model update transmission and edge-assisted verification (Hakeem et al.
, 2.
This extension would require careful co-design of learning algorithms and communication protocols to meet stringent reliability requirements.
Conclusion This study articulates a scalable pathway for reconciling the often-conflicting objectives of data privacy and operational efficiency in sustainable digital logistics.
By synthesizing local outlier filtering, sparse ternary compression, and differential privacy into a unified PrivacyPreserving V2X Federated Learning (PPVFL) framework, we have successfully demonstrated that preserving digital sovereignty need not come at the expense of model utility.
Empirical validation through a pilot deployment in Hanoi confirms that delegating quality control to the network edge significantly mitigates the noise inherent in vehicular sensor data, thereby outperforming conventional federated approaches in both prediction accuracy and communication overhead.
These findings directly address the critical challenges of adapting digital solutions to the dynamic traffic patterns of developing economies, a gap previously highlighted in logistics literature (Wang et al.
, 2023.
Tran, 2.
Furthermore, the frameworkAos seamless integration with existing management systems offers a practical mechanism for logistics operators to comply with VietnamAos stringent data localization laws while advancing carbon reduction targets (Murphy & Poist, 2.
Looking forward, we envision extending this architecture to support latency-sensitive applications within emerging Privacy-Preserving V2X Federated Learning with Local Filtering for Sustainable Digital Last-mile Logistics 6G ecosystems (Adhikari & Hazra, 2.
, specifically exploring adaptive privacy mechanisms that can dynamically respond to real-time safety-critical scenarios (Hakeem et al.
, 2.
This work positions privacy-aware federated learning not merely as a compliance tool, but as a foundational enabler for the next generation of green, intelligent urban transport systems.
References