Cross-border logistics risk warning system based on federated learning

Abstract

As international trade grows, managing cross-border logistics becomes more complicated and riskier. This paper aims to build a secure and private risk warning system that allows different logistics partners to collaborate without sharing sensitive information. Traditional systems that try to predict these risks collect all data in one place, creating privacy issues and making it hard for companies and governments in different regions to work together. This paper proposes a framework called the Secure and Federated Logistics Risk Warning System using Federated Learning (SafeLogFL). SafeLogFL is a privacy-preserving risk alert system for cross-border logistics that ensures secure, decentralized collaboration across different entities involved in logistics. Instead of sharing sensitive data, each participant trains the model locally on its data. The locally trained model updates are then aggregated using the Federated Averaging (Fed Avg) algorithm, which ensures model convergence while maintaining data privacy. Experimental results showed an average accuracy of 91.3% and compliance with privacy laws such as the General Data Protection Regulation (GDPR), proving the system’s effectiveness in predicting delays, disruptions, and compliance issues. SafeLogFL provides a scalable, privacy-preserving solution for managing risks in global logistics, fostering secure collaboration between multiple parties.

Introduction

The complexity of cross-border logistics in global trade

International trade is the pillar of the contemporary economy, with commodities crossing continents daily to reach businesses and consumers¹. International trade entails complex logistics across air, sea, and land modes and involves seamless coordination among varied stakeholders in different parts of the world². With increased volumes of trade, so does the complexity of cross-border logistics management. Such systems must arrange an intricate web of transporters, customs agents, transport companies, port administrators, and regulatory agencies operating within a range of national policies and international agreements³. Based on the destination and the origin, each cargo needs to overcome a series of logistical hurdles and administrative procedures that can be highly dissimilar across jurisdictions⁴.

This logistical heterogeneity raises the probability of delays and errors, particularly when handling non-standardized documents or protocols that evolve rapidly⁵. Trade facilitation is also hindered by time zones, language differences, and coordination issues of technological capabilities among international partners⁶. Therefore, Global logistics players call for more interoperable, intelligent systems that can simplify coordination and reduce border friction⁷. Figure 1 shows the risk factors in cross-border logistics networks.

Fig. 1.

Risk factors in cross-border logistics networks.

Artificial intelligence in logistics

Artificial Intelligence (AI) application in logistics transforms how operations and risks are handled. AI models can process huge amounts of structured and unstructured data, recognize patterns, and anticipate disruptions before a disruption occurs⁸. Machine learning algorithms are used to route optimize, warehouse performance, make delivery-time predictions, and identify anomalies indicative of fraud or regulatory noncompliance⁹. AI can revolutionize logistics planning, execution, and surveillance, provided it receives quality data¹⁰.

In cross-border logistics, AI facilitates preventive measures against risks proactively by scanning for vulnerabilities in the supply chain and proposing prophylactic interventions¹¹. Predictive analysis powered by AI can prevent delays at customs, detect discrepancies in shipping documents, or forecast regulatory shifts that impact cargo movement¹². The increasing integration of AI technology into logistics infrastructure provides real-time decision support that can sharply minimize downtime, cost, and error. By employing federated learning, AI can achieve its full potential without violating the sovereignty and privacy of the data contributors¹³ (Table 1).

Table 1.

Summary of the problem statement.

Problem	Centralized systems	Decentralized framework (SafeLogFL)
Shipment delays	Slow detection due to centralized data processing	Faster detection with real-time, local updates
Regulatory changes	Hard to adapt quickly; needs central system updates	Quick local adaptations to new regulations
Fraud detection	Limited visibility; relies on delayed or incomplete data	Broader, real-time insights from multiple participants
Data privacy	Low—sensitive data collected centrally, risk of breaches	High—data stays local, minimizing exposure
Collaboration	Difficult—low trust among stakeholders due to privacy concerns	Easier—privacy-preserving participation builds trust
System vulnerability	High risk—single point of failure	Low risk—distributed, resilient structure

Contributions of the paper

• To introduce SafeLogFL, a privacy-preserving risk warning system for global logistics using federated learning.

• To implement a decentralized training framework leveraging MLP and Fed Avg to avoid centralized data sharing.

• To validate the system’s effectiveness with experimental results showing over 91% accuracy and GDPR compliance.

• To foster secure, scalable stakeholder collaboration across diverse international logistics environments.

Related works

Cross-border logistics has become more complicated and vulnerable to delays, regulations, and geopolitical intervention threats. Conventional risk management depends on using centralized data, which contradicts data privacy, stakeholder trust, and regulatory compliance, such as GDPR. Such systems are opposed due to the need to reveal confidential information. FL seems like an answer that facilitates decentralized model training without revealing original data; hence, confidentiality is maintained. This makes FL very apt for cross-border collaborative risk forecasting in global logistics. The article discusses prevailing risk management measures, data exchange issues, and how FL fills gaps in cross-border logistics systems.

Traditional approaches to risk management in cross-border logistics

Legacy risk management systems used in cross-border logistics are typically based on centralized data collection, wherein information from various sources is aggregated and analyzed at one location. These systems are mainly based on historical trends and rule-based models to measure risk. They have limited real-time flexibility, have resistance from stakeholders because of privacy issues, and offer higher exposure to data breaches, all of which reduce their relevance to the existing fast-paced and globalized logistics scenario. Table 2 summarizes Traditional Approaches to Risk Management in Cross-Border Logistics.

Table 2.

Summary of traditional approaches to risk management.

Author(s)	Proposed technology	Method used	Result	Advantage	Limitation	Research gap
Almusawi & Pugazhenthi14	Multi-agent supply chain scheduling	Honey bee optimization	Optimized resource allocation in supply chains	Bio-inspired optimization improves efficiency	May not scale for very large supply chains	Limited integration with real-time logistics data
Zhou et al.15	Risk aversion in B2C cross-border E-commerce	Quantitative risk analysis	Identified risk factors affecting B2C cross-border trade	Highlights key risk factors	Focused only on B2C, ignores B2B dynamics	Need for predictive risk models in real-time
Dai & Min16	Cross-border E-commerce Reform	Quasi-natural experiment	Reform reduced supply chain risks	Provides empirical evidence	Regional pilot zone, limited generalizability	Lack of predictive mechanism for risk mitigation
Zhang et al.17	Supply Chain Resilience vs. Vulnerability	Analytical modeling	Balances resilience and vulnerability	Provides framework for decision making	Limited to modeling; no real-time application	Real-time predictive tools for resilience needed
Brookbanks & Parry18	Industry 4.0 in Supply Chains	Case study	Improved resilience of established supply chains	Highlights Industry 4.0 benefits	Case-specific results	Integration with cross-border e-commerce not fully explored
Hertzel et al.19	Global Supply Chains & Financing	Review & analysis	Identified financing risks	Comprehensive overview	Focused on financial aspects	Predictive risk mitigation models missing
Dawar & Bai20	Supply Chain Risk Management	Literature review	Summarized existing risk management approaches	Provides holistic understanding	No empirical evaluation	Need for AI-enabled risk prediction systems
Settembre-Blundo et al.21	Sustainability-based Risk Management	Decision-making framework	Enhanced corporate flexibility and resilience	Sustainability integrated into risk management	High-level conceptual	Application in cross-border logistics not explored
Teramura & Shimatani22	Ecosystem-based Disaster Risk Reduction	Case study	Demonstrated environmental benefits	Promotes eco-friendly solutions	Context-specific (Japanese river)	Application to industrial supply chains lacking
Sahab et al.23	Agroecosystem Risk Assessment	Environmental modeling	Assessed soil salinity risks	Provides scientific risk assessment	Limited to agricultural domain	Integration with economic and logistic factors
Olaniyi et al.24	CyberFusion Protocols	Enterprise risk management + ISO 27,001 + mobile forensics	Enhanced digital security in enterprises	Combines multiple security standards	Complex implementation	Scalability and cross-border compliance not addressed
Benami et al.25	Agricultural Risk Management	Remote sensing + Crop modeling + Economics	Optimized agricultural risk mitigation	Multi-disciplinary integration	Requires extensive data	Limited to agriculture; cross-sector application missing
Lazaros et al.26	Federated Learning	Survey & analysis	Reviewed collaborative intelligence techniques	Comprehensive review	Theoretical; no empirical evaluation	Benchmarking real-world applications required
Farahani & Monsefi27	Smart & Collaborative Industrial IoT	Federated Learning + Data Spaces	Enabled secure collaborative IoT analytics	Privacy-preserving approach	Industrial IoT only	Scalability in large networks untested
Sharma et al.28	Healthcare Federated Learning	Collaborative AI	Improved healthcare predictions without data sharing	Ensures patient privacy	Focused on healthcare domain	Adaptation to logistics/cross-border trade missing
Gong et al.29	Collaborative Intelligence Systems	Systematic review	Identified AI collaboration methods	Guides future research	Lacks implementation insights	Need for domain-specific deployment
Nguyen et al.30	Smart Healthcare via FL	Survey	Comprehensive overview of FL applications	Detailed FL techniques and challenges	Healthcare-focused	Real-time predictive risk management in logistics unexplored
Mastoi et al.31	Explainable AI for Brain Tumor	Federated Learning + XAI	Interpretable medical predictions	Transparency in AI decisions	Healthcare-specific	Transferability to logistics risk prediction needed
Hu, N.32	Fuzzy Neural Network	Risk assessment modeling	Improved accuracy in assessing logistics risks in agricultural e-commerce	Handles uncertainty in data effectively	Specific to agricultural products	Does not address real-time decentralized data or privacy concerns
Mu, W.33	Warning Model (Not specified)	Statistical analysis and risk modeling	Identified major risk sources in cross-border e-commerce logistics	Practical model for early warning	Lack of technological specificity and real-time applicability	No integration of privacy-preserving models
Jing & Yang34	Big Data Technology	Data-driven risk prevention	Enhanced risk prediction via big data analytics	Leverages large data volumes for better forecasting	Centralized data processing	Lacks privacy preservation and stakeholder trust mechanisms
Ylönen & Aven35	Intelligence Integration Framework	Conceptual analysis in border control context	New approach to integrating risk intelligence	Focused on customs and regulatory risk	Limited practical application	Does not incorporate machine learning or decentralized techniques
Yan et al.36	Third-Party Logistics Risk Analysis	Analytical hierarchy process (AHP)	Identified criteria for logistics partner selection	Structured decision-making tool	Limited to partner selection	Ignores operational and real-time risk predictions
Bandaranayake et al.37	Reference Model	Model development and validation	Improved logistics performance through structured modeling	Comprehensive analysis for process improvement	Not focused on risk or AI-based systems	Lacks integration with predictive technologies
Kalogiannidis38	Disaster Response Analysis	Qualitative risk analysis	Identified key cross-border coordination challenges	Provides insights into disaster logistics	Not specific to e-commerce or FL	No technological implementation or predictive modeling discussed

Data privacy challenges and the need for decentralized collaboration

Global logistics risk management entails working with other individuals’ confidential information. In the present era, with strongly regulated data security legislation such as GDPR, organizations would not wish to exchange proprietary information through centralized services. This hinders data exchanges and, in turn, precision in risk forecast models. These problems can be solved by decoupling so that cooperation will be attained without undermining confidentiality, developing confidence, and creating a more widespread and stronger risk management framework. Table 3 summarizes the Data Privacy Challenges and the Need for Decentralized Collaboration.

Table 3.

Summary of the data privacy challenges and the need for decentralized collaboration.

Author(s)	Proposed technology	Method used	Result	Advantage	Limitation	Research gap
Bolatbekkyzy, G.39	Legal Framework for Digital Government	Legal analysis	Identified key legal barriers to cross-border data transfer	Raises awareness of regulatory challenges in digital governance	Focused more on legal than technical aspects	Lacks technological solutions to data transfer issues in logistics
Supangkat et al.40	Cross-Border Digital Identity Systems	Systematic analysis and policy evaluation	Highlighted implementation challenges of digital identity for public infrastructure	Addresses cross-border trust and security in digital systems	Limited focus on logistics-specific applications	Does not integrate with logistics data-sharing or risk prediction frameworks
Cecil, P.41	Supply Chain Optimization Strategies	Strategic framework analysis	Offered guidelines for efficient cross-border operations	Balances speed and cost in supply chain design	Lacks data-driven risk prediction tools	No focus on AI or privacy-preserving data use
Cao, J.42	Blockchain-based Data Management	Device and system evaluation	Improved reliability and traceability in e-commerce data management	Enhances data integrity and transparency	Complexity in implementation and scaling	Not integrated with federated learning or real-time predictive models
Sun & Gu43	IoT-based Supervision System	System optimization approach	Enhanced logistics supervision and visibility	Real-time tracking and automation	Potential security and scalability concerns	Does not address multi-party privacy or predictive modeling
Zhang, Xu & Xiao44	Big Data Decision System with Data Fusion	Intelligent data fusion techniques	Improved decision-making through integrated data sources	Supports intelligent perception and faster response	Data centralization may raise privacy concerns	Lacks decentralized and privacy-preserving model frameworks
J. Li45	Federated Learning for Address Classification	Address classification for delivery optimization	Improved classification accuracy while preserving local data privacy	Preserves data privacy, enhances delivery accuracy	Focused only on address verification; not cross-border	No cross-border, no risk prediction, no real-time alerts
M. Chen46	Federated Learning for Logistics Data Sharing	Secure multi-party logistics data sharing	Enhanced data security and enabled collaborative learning without centralizing sensitive information	Secure collaborative learning, preserves data privacy	Does not include predictive risk modeling, cross-border insights, or stakeholder-specific alerts	No risk modeling, no cross-border insights, no regulatory compliance
G. Zheng, L. Kong, A. Brintrup47	FL for supply chain risk prediction	Federated Learning	Effective collaborative risk prediction without sharing raw data	Preserves privacy, supports distributed collaboration	Focused on generic supply chain data; no real-time or cross-border adaptation	Does not address non-IID data or domain-specific metrics
A. Brintrup48	GNN + FL for supply chain visibility	GNN + FL	Improved prediction and visibility across decentralized supply chains	Integrates advanced analytics with privacy-preserving FL	Computationally intensive	Limited focus on adaptive updates and stakeholder collaboration in real-time cross-border logistics

Table 3 shows that Li’s preliminary research places a strong emphasis on verifying addresses and optimizing delivery processes by utilizing e-commerce shipment and customer data. On the other hand, SafeLogFL supports a large number of stakeholders, predicts risk across boundaries, and dynamically adjusts logistics scenarios to accommodate modifications. It does not address border supply chain risk prediction, adaptable updates to models, or real-time alerts. M. Chen emphasizes the secure exchange of data across logistics partners through the utilization of multi-party operational data. However, the study does not include prediction risk analysis, cross-border operating insights, stakeholder-specific alerts, or compliance with GDPR. Federated learning is utilized by SafeLogFL to ensure compliance with regulatory requirements, predict risks while protecting users’ privacy, and make real-time decisions across global logistics networks.

This research effort expands upon the ideas that were described in^47,48, which illustrate how Federated Learning (FL) makes it possible to train models in dispersed environments while maintaining the confidentiality of the data. By extending these ideas, SafeLogFL is able to particularly solve difficulties that are associated with cross-border logistics. It manages non-iid data partitioning between partners, reflecting real-world variances in warehouses, shipping routes, and cargo categories. This is in contrast to usual FL implementations in the healthcare or Internet of Things industries. In addition to this, the system features adaptive model updates that are designed to react to changes in operational conditions, such as delays, differences in the weather, or policy modifications. The Scalability Score (SS) and the Stakeholder Teamwork Index (SCI) are two additional metrics that SafeLogFL has introduced in order to assess the effectiveness of the system and the level of collaboration among the various stakeholders.

Research problem

Historical trends are enlightening, but data privacy and efficiency in operation get compromised. Stakeholders are averse to centralized methods as sensitive business information is at stake, along with the problem of complying with regulations, particularly in severe rules like the GDPR. Centralized methods also risk bottlenecks and susceptibilities that keep them from making dynamic changes. There is an immediate need for such a risk management model to facilitate privacy protection, decentralization, and data sharing with better prediction ability. Such a model allows different entities, e.g., shipping operators, customs clearing authorities, port operators, etc., to interoperate without revealing their confidential data while establishing confidence and cooperation everywhere in the world in the networked supply chains. Table 1 illustrates the problem statement summary.

SafeLogFL methodology

SafeLogFL is a secure and Federated Logistics Risk Alert System that uses FL for privacy-aware cross-border logistics risk warning collaboration. It maintains confidentiality by training local models on confidential datasets, communicating model updates to a central server, and computing them with the Fed Avg algorithm. Figure 2 shows the working process of the SafeLogFL framework.

Fig. 2.

Workflow of the SafeLogFL framework.

Data collection

Data collection in the SafeLogFL framework is based on a decentralized architecture where each party e.g., shipping lines, customs, and port operators has and processes its operational data locally. This strategy protects sensitive data within organizations, enabling GDPR compliance and data privacy. Each organization warehouses structured records like shipping information customs information, and disruption to operations. SafeLogFL supports local data management and computation to prevent data exposure and enable secure interactive instruction without central hosting.

A non-IID partitioning technique was used to distribute the dataset across the clients in the SafeLogFL studies. This was done to provide an accurate representation of the cross-border logistics scenarios encountered. Only the local operational data of each client is stored, and each client corresponds to a stakeholder, such as a shipping business, port authority, or customs agency. The data segmentation was carried out according to geographical locations and the types of shipments, which meant that each customer had records from specific warehouses, shipping routes, or cargo classifications. This guarantees that local models capture patterns that are distinct to stakeholders, thereby emulating the heterogeneity that exists in the real world. The FedAvg algorithm aggregates the locally trained models, taking into consideration the non-IID distribution, while also protecting the confidentiality of the data. Due to its close simulation of genuine multi-stakeholder logistics networks, this partitioning technique enhances the validity and reproducibility of the results.

Data preprocessing

Every stakeholder () in the SafeLogFL platform does local preprocessing on its dataset () to prepare the data for model training. The process starts with feature extraction, where appropriate attributes are chosen or extracted from raw logistic data. Such features include shipment weight, transport distance, port traffic index, and history of delays records, mathematically represented as a feature vector . To ensure numerical stability and boost the performance of the model, each stakeholder standardizes its data individually using the formula ​​ ​, in which and ​ are the mean and standard deviation of the features of dataset (). Subsequently, each data instance is assigned a label , based on historical results. For binary risk classification, the label is denoted as in Eq. 1.

1

Preprocessing operations are done locally so that no raw data is communicated outside the organization. Model parameters (e.g., weight updates) are only passed during federated training time, shown as .

Model training with federated learning

In the SafeLogFL framework, after local data preprocessing, every involved stakeholder learns its machine learning model independently on its local data, e.g., shipping company, port authority, or customs agency. This regional model does not share sensitive information and is a Federated Learning (FL) building block.

Model architecture: Multi-Layer perceptron (MLP)

In SafeLogFL, a neural network MLP learns from structured log data like shipping data, port traffic, and customs manifests. Figure 3 shows the architecture of an MLP in the SafeLogFL framework.

Fig. 3.

Architecture of an MLP In the SafeLogFL Framework.

MLP is an artificial feed forward neural network that receives input features and outputs predictions, propagating through a series of computation layers. It can be applied particularly well to tabular classification problems, e.g., forecasting delay or logistics disruption. The structure of an MLP with () layers are shown in Table 4.

Table 4.

Structure of the MLP model.

Layer	Purpose	Operation/formula
1. Input Layer	Receives the input feature vector , where d is the number of features.	No computation; forwards the input to the next layer.
2. Hidden Layers	Core computation: learn complex patterns using linear transformation and activation functions.
3. Output Layer	Generates the final prediction

Where ​ is the output from the previous layer (or input ) ​ is the weight matrix. ​ is the bias vector, is the activation function (typically ReLU: ), is the sigmoid function is for binary classification, ​ are the weights and biases of the output layer, ​ is the output of the last hidden layer, represents the predicted probability of a shipment being delayed. Equation 2 summarizes the overall output of the MLP with layers.

2

This expression denotes the transformation of the input through a series of layers, each applying weighted sums and activation functions, ultimately producing a probabilistic output. Figure 4 shows the MLP layer flow for logistics in the SafeLogFL framework.

Fig. 4.

MLP layer flow in logistics.

Model updates and global model aggregation in federated learning

After the model has been initialized, every participant (e.g., a warehouse or a supplier) conducts local training on its data and calculates model updates, i.e., weights and gradients. These updates, not the original data, are uploaded to the central server. The central server collects these updates via the Federated Averaging (Fed Avg) algorithm and averages all participants’ weights to create an accurate global model without sacrificing privacy.

The global model aggregating process enables the central server to combine the locally updated models into a global model. The updated global model signifies learning from data for all stakeholders, acquiring valuable insights unavailable in any individual dataset. The global model is remitted to participants after aggregation, updated with respective data if required.

1. Initialization: The central server initializes a global model ​.

2. Broadcast: The server sends the current global model ​ to selected clients (e.g., individual warehouses).

3. Local Training: Each client trains the model on its local data for a few epochs and computes updated model parameters ​.

4. Send Updates: Each client sends its updated model ​ back to the central server.

5. Aggregation (Fed Avg): The server aggregates all client updates using Fed Avg to obtain the new global model ​.

6. Repeat: this process continues for multiple rounds.

Algorithm 1 describes training a global model through FL in order. It preserves data privacy because clients can train models locally and transmit only model updates to a central server, aggregating them using the Fed Avg technique.

Algorithm 1.

Federated averaging (FedAvg) for privacy-preserving collaborative model training in distributed systems.

In a FL environment, the training process starts with the central server starting the weights of the global model as . Every client () keeps its local data set represented as (), which is kept on the client to maintain data privacy. For every communication round , a subset of clients () is picked to help train. The global model is then sent to these selected clients. Each client trains locally on its data () using the global weights as initial weights. The process, Local Train (​ ,​ ), produces new local model weights (). Apart from these weights, each client sends the local dataset size, . The global server then aggregates all client updates and calculates the new global model () based on the Fed Avg algorithm. This is done by calculating a weighted average of the clients’ updates to their models, where each client’s contribution to the aggregate is weighted by its data size compared to the total data in all the clients involved. Figure 5 shows the global model aggregation in federated learning.

Fig. 5.

Global model aggregation in federated learning.

Every client , one of the different cross-border logistics participants, trains the model locally on its local dataset and receives an updated model . The number of samples from the data in client k is . Once locally trained, the clients return their new model weights to the server. The server then aggregates the global model and the Fed Avg algorithm. The updated global model is calculated as a weighted sum of the clients’ updates, with each client’s update weighted by the size of the local data, as in Eq. 3.

3

Here, stands for the total number of clients’ samples, and for the number of clients encompassed. During the aggregation phase, clients with more or broader data sets would have greater control over the novel global model.

The aggregate ​model captures are then redistributed to all participants to continue training or fine-tuning. This continuous process improves the model’s risk factor capture capacity and anomaly prediction in various logistics environments, yielding a strong cross-border risk warning system that ensures data sovereignty and privacy regulations.

Model update frequency and adaptability

The update frequency of how SafeLogFL updates its model plays a crucial role in maintaining the system’s accuracy and responsiveness to new threats. Depending on circumstances, the model updates might be in varied forms:

• For operations or routes that do not often alter, updates can be batched—e.g., once a day or once weekly. This is suitable if fast change is less likely to occur and the compute and network usage is less.

• More updates can be done in rapidly changing or high-risk situations. Here, everyone often updates the model on small chunks of new information and transmits more quickly. This allows the system to respond more to new problems such as delays, weather, or immediate policy changes.

SafeLogFL’s adaptive nature enables it to work efficiently even in fluctuating scenarios, updating the global model accurately and consistently. Table 5 compares SafeLogFL’s model update strategies based on operational context, responsiveness, and system efficiency.

Table 5.

SafeLogFL model update modes and their use cases.

Update mode	Description	When to use	Advantages
Batch Update	Model updates sent periodically (e.g., daily or weekly)	Stable environments with fewer sudden changes	Lower communication cost, simple setup
Frequent/Online Update	Model updates sent more frequently after small local training rounds	Dynamic environments with frequent disruptions	Higher responsiveness, real-time adaptation
Hybrid Approach	Combines batch and online updates based on context	Mixed-risk routes or flexible stakeholder needs	Balanced performance and efficiency

Risk prediction and alerts using the federated global model

After the global model has been trained and retrained using the Federated Learning (FL) process, it is applied in carrying out risk forecasting operations on the logistics network. The key aim of this phase is early identification of potential disruption in advance and issuing warnings to stakeholders like warehouses, suppliers, and transportation companies. Cross-border logistics shared risk drivers are Ship-out delays because of border congestion or customs inspection, Customs issues because of incomplete or conflicting documentation, Supply chain disruption because of weather, labor strikes, or inventory differences. With the trained global model, every participating client can forecast whether an impending shipment or operation will likely be at risk, based on features derived from historical and real-time logistics data.

Input features for risk prediction

Let ​ represent the feature vector for the shipment or transaction. Each sample \mathbf{x}_i​ is associated with a label , as in Eq. 4.

4

Prediction using the global model

Let Denote the trained global model parameterized by weights . Given an input feature vector ​The model outputs a predicted probability of risk. This can be obtained through Eq. 5.

5

If , where is a predefined risk threshold, then the shipment is considered at risk and an alert is triggered. The alert mechanism is shown in Eq. 6.

6

The threshold The threshold may be set based on the stakeholder’s risk tolerance. For example, it may be lower for high-priority global shipments to detect more potential anomalies, whereas it may be higher for regular shipments to reduce false alarms. If.

Feeding this into the global model yields,

.

If τ = 0.75, then .

Therefore, the system gives the logistics operator an early warning to react, such as rerouting, adding documentation, or pre-notifying customs.

SafeLogFL can be more responsive by integrating it with real-time logistics tracking systems and live data feeds. By being linked to live data feeds, such as shipping tracking systems, GPS data, customs inspection status, or weather forecasts, SafeLogFL can be fed continuous new input features for risk prediction. The integration enables proactive decision-making, and logistics partners can take early action to preclude or reduce delays and compliance problems.

Privacy and compliance in federated learning

The use of FL on a cross-border logistics warning system is its capacity to keep sensitive and proprietary information secure while allowing for concurrent model training and cooperative sharing. In FL, unprocessed training data is never disclosed beyond the organizational boundary or the local device. All users update the model locally based on their local dataset (operational records, shipment files, and history of incidents, e.g.) and send just the model updates to the server (as weights or gradients, say). Nothing that involves revealing a single point data is shared from any update.

Legal and regulatory compliance

Federated Learning supports data protection laws by design directly:

GDPR: GDPR is focused on data minimization and locality. FL aligns with these principles since it prevents the transfer or centralization of personal data.

Data Residency Requirements: Certain jurisdictions mandate that data created within their borders be stored and processed locally. FL ensures computation is done where the data is located, inherently satisfying such regulations.

Confidential Business Information: Global supply chain firms view logistics information as confidential. FL architecture allows for cooperative intelligence without any party revealing sensitive operation details.

Security improvements

Apart from imposing compliance and privacy, FL systems can also include:

Secure Aggregation Protocols: These protocols do not allow the server to observe a client update, only the aggregated result.

Differential Privacy: By introducing controlled noise to the updates, FL systems can make individual data records impossible to infer even from model updates.

Authentication and Access Control: Only authorized participants can engage in training rounds, excluding unauthorized access to data exchanges and model updates.

Collaborative trust model

The decentralized nature of FL facilitates a secure environment where autonomous stakeholders with varying legal and cultural paradigms can yet find value in shared accumulated knowledge. This is particularly relevant in cross-border logistics where international coordination is necessary, but information sharing has legal and ethical limits.

Table 6 compares typical centralised logistics systems with the SafeLogFL solution. Typical systems gather all data in one location, which means greater risks of data breaches and hinders compliance with privacy regulations. SafeLogFL has a decentralised system where data is on the periphery and only transmits model updates. This significantly minimizes privacy attacks, simplifies GDPR compliance, and allows secure regional collaboration. Therefore, SafeLogFL presents a safer and privacy-conscious solution for risk management of worldwide logistics.

Table 6.

Comparison of traditional systems and safelogfl in terms of privacy and collaboration.

Aspect	Traditional system	SafeLogFL approach
Data Centralization	Yes	No (decentralized)
Risk of Data Leakage	High	Very low
GDPR Compliance Difficulty	High	Easier to achieve
Global Collaboration	Hard (due to legal barriers)	Smooth and secure

Research methodology

This paper introduces the Secure and Federated Logistics Risk Warning System with Federated Learning (SafeLogFL) to address these issues. SafeLogFL utilizes a Multi-Layer Perceptron (MLP) neural network to learn the structured logistics data like shipment history and customs declarations. All parties train the model locally using their local data without sharing any raw data. Model updates are exchanged and combined by using the Federated Averaging (Fed Avg) algorithm. This decentralized training approach maintains confidentiality, honors regulatory compliance, such as GDPR, and increases system resilience by eliminating points of central failure. This approach facilitates cooperative risk forecasting among entities while safeguarding sensitive data and provides real-time alerting for cross-border logistics delay, fraud, and non-compliance.

Dataset explanation

The Logistics and Supply Chain Dataset posted on Kaggle⁴⁵ includes logistics and supply chain activity data. The dataset was sampled from a Southern California logistics network, thus providing real-life data about logistics operations. The dataset includes features like order dates, shipping time, delivery dates, transit times, warehouse and customer locations, shipping modes, freight cost, and product categories. Such functionalities allow for the simulation and analysis of logistics operations, such as delivery performance and risk conditions. The data set is well-suited for training machine learning models capable of creating risk labels (e.g., on-time versus delayed delivery) and emulating decentralized data scenarios using region or logistics hub partitioning. It is well-suited for experimentation with and developing privacy-enabling federated learning models such as SafeLogFL.

Table 7 classifies the major characteristics of the SafeLogFL system for predicting global logistics risk. Characteristics include geographic information, shipment information, customs information, risk factors, operation parameters, training labels for models, and privacy controls. The organization encourages thorough risk estimation while ensuring data privacy via federated learning. All groups enable various delay, disruption, and compliance risk detection aspects, including foreign logistics activities.

Table 7.

Categorization of cross-border logistics risk-related features used in the safelogfl system.

Category	Feature examples	Purpose
Geographic Info	Origin Country, Destination Country	Understand region-specific risks, regulations, and geopolitical factors
Shipment Details	Shipment Mode, Departure Date, Expected Arrival, Actual Arrival	Detect delays, mode-specific vulnerabilities
Customs Data	Clearance Time, Inspection Status, Documentation Completeness	Assess the likelihood of customs delays or compliance issues
Risk Indicators	Fraud History, Trade Regulation Changes, Tariff Change Flags	Anticipate possible disruptions or fraudulent activities
Operational Metrics	Shipment Cost, Handling Time, Storage Duration	Monitor for inefficiencies and bottlenecks
Model Training Labels	Delay Flag, Compliance Issue Flag	Supervised learning outcomes (for prediction and classification)

Figure 6 shows the use of federated learning to predict cross-border logistics risk. The various entities, such as shipping companies and customs, participate in the architecture. Participants train their local MLP model using their respective logistics data, e.g., shipment history and clearance records. Instead of exchanging sensitive information, participants update models and send them to a central aggregator, which aggregates them based on the Fed Avg algorithm to generate a global risk prediction model. The generated model is then shared with all the participants so there can be ongoing learning without violating data privacy. The method ensures collaborative and secure risk prediction among stakeholders without violating data protection laws such as GDPR.

Fig. 6.

Decentralized training process of SafeLogFL.

Experimental setup

To assess the performance of the proposed SafeLogFL framework, experiments were conducted using the publicly available Logistics and Supply Chain Dataset from Kaggle, which contains structured data on shipments, customs status, delivery performance, and warehouse operations. SafeLogFL was compared against three established risk management methods: the Fuzzy Neural Network Risk Assessment Model³², The administration of supply chain data can be carried out in a secure and decentralized manner with the help of this framework, as detailed in⁴². The structure, which guarantees the data’s integrity, transparency, and traceability, constitutes a solid foundation for the deployment of predictive models. This foundation is built by the structure. With the help of this framework, the exploitation of the predictive method makes it possible to anticipate risks with more precision. Additionally, it highlights how the combination of blockchain-based data management and advanced analytics may improve the efficiency of supply chain operations. the Blockchain-Based Data Management Device, and the IoT-Based Logistics Supervision System⁴³. These methods represent a range of traditional and modern logistics management approaches. Evaluation metrics included prediction accuracy, data privacy and security compliance, scalability and stakeholder collaboration, and computational efficiency, reflecting technical capabilities and legal requirements for cross-border logistics systems.

The SafeLogFL model was implemented using a Multi-Layer Perceptron (MLP) neural network trained with Federated Learning across ten simulated logistics clients, such as warehouses and customs checkpoints. Each client retained its private dataset in a non-identically distributed (non-IID) format to replicate real-world data silos. During training, clients performed local model updates and sent only the encrypted weights to a central server, which applied the Federated Averaging (Fed Avg) algorithm to compute the updated global model. The system completed 100 training rounds, with each client running multiple local epochs per round. Secure aggregation and differential privacy mechanisms were used to ensure compliance with international data protection laws, including GDPR. The entire setup was built using TensorFlow Federated and PySyft, and deployed on a distributed virtual infrastructure. This configuration validated SafeLogFL’s ability to deliver accurate predictions while preserving privacy, ensuring scalability, and supporting decentralized collaboration.

The data silos that exist in the real world of logistics result in datasets being distributed among clients in a manner that is not IID. It is ensured that each client holds heterogeneous but realistic parts of the overall dataset by performing partitioning depending on the location of the warehouse and the shipping routes. With a total of 128, 64, and 32 neurons, the MLP in SafeLogFL is comprised of three hidden layers. ReLU activation is used for all the hidden layers, while sigmoid activation is used for the output layer, which is responsible for binary risk prediction. To ensure reproducibility and consistent performance under federated learning, the model is trained using the Adam optimizer with a learning rate of 0.001, a batch size of 64, and ten local epochs for each client during each communication round during the training process.

Experimental analysis and results

Prediction accuracy

Prediction performance is a key measurement to check the degree to which SafeLogFL identifies prospective logistics risks like delay in shipment, customs clearance infringement, and route disruption. Running in a federated environment, the prediction task is a binary or multi-class classification model over distributed non-IID datasets without aggregating raw data. The global model, which is then received by federated averaging, is then evaluated using a centralized test set to quantify performance.

Let the model’s prediction function be denoted by , where = the parameters obtained through federated learning (Fed Avg), and = the input feature vector (e.g., shipment size, route history, customs code). The accuracy can be mathematically expressed as in Eq. 7.

7

where is the total number of samples in the test dataset. ​ is the true class label for sample . , is the predicted probability distribution over the classes. gives the predicted class. is the indicator function that returns 1 if the prediction matches the true label, 0 otherwise.

SafeLogFL, Fuzzy Neural Network, Blockchain-Based System, and IoT-Based System logistics risk prediction accuracy are shown in Fig. 7. Each plot indicates SafeLogFL’s best risk category model accuracy. Visual comparison reveals that federated learning improves prediction accuracy in all circumstances without compromising compliance or scalability, making it helpful in cross-border logistics. Figure 7(a) shows how accurately all models anticipate logistical concerns. SafetyLogFL has the highest accuracy (91.3%), indicating dataset generalization. Blockchain-based and fuzzy neural network models degrade rapidly. The Internet of Things behaves poorly. Federated learning can learn from distributed patterns without compromising data privacy, proving SafeLogFL’s excellence. Figure 7(b) compares all models’ customs delay and infraction detection. SafeLogFL handles sensitive regulatory data at 92.1%. Second, fuzzy neural networks, followed by blockchain and IoT. Localised training using region-specific customs records without data centralization makes SafeLogFL excellent for cross-border regulatory risk management. Figure 7(c) shows how effectively each model forecasts shipment delays, a significant logistics issue. SafeLogFL again scores 90.5% accuracy across datasets, demonstrating delivery timetable awareness. Fuzzy Neural Networks are slightly less accurate than Blockchain and IoT. While retaining privacy and logistical independence, federated learning allows stakeholders to forecast delays. Models for transport route disruption are shown in Fig. 7(d). SafeLogFL performs best at 91.0%, including local knowledge without sacrificing routing-sensitive information. Transit information discrepancies weaken IoT-based systems, whereas Fuzzy Neural Network and Blockchain-based models are good. SafeLogFL thrives in global supply networks’ dynamic and uncertain routes.

Fig. 7.

Accuracy comparison analysis (a) Overall prediction accuracy, (b) Customs Risk Prediction accuracy, (c) Shipment Delay Prediction Accuracy, and (d) Route Disruption Prediction accuracy.

Data privacy and security compliance

Data Privacy and Security Compliance in federated learning for cross-border logistics risk warning systems refers to how well the system ensures the data’s confidentiality, integrity, and availability while adhering to legal and ethical standards (e.g., GDPR, CCPA). In federated learning, data privacy is ensured by performing computations locally and only sharing model updates rather than raw data. DPC is a function of various privacy measures, represented by Eq. 8.

8

where Encryption refers to the level of encryption used to protect data. Anonymization refers to the degree to which sensitive data is anonymized. Data Sharing Protocol refers to securely sharing model updates (e.g., secure multi-party computation, differential privacy). Regulatory Compliance is the degree to which the system adheres to privacy laws and regulations. Auditability is the ability to track and verify data access and usage. The security function is shown in Eq. 9.

9

where Data Integrity refers to ensuring that data hasn’t been tampered with. Risk Exposure is the probability that data is exposed to risks (e.g., cyberattacks, leaks). Security Measures refer to tools like encryption, firewalls, or intrusion detection systems.

Figure 8 contrasts the cross-border logistics risk warning system model of SafeLogFL, Fuzzy Neural Network (Fuzzy NN), Blockchain-Based System, and IoT-Based System regarding Privacy, Security, and Regulatory Compliance. SafeLogFL ranks higher than the rest consistently on every aspect, with privacy being 95%, security being 90%, and regulatory compliance being 92%. These are high due to its decentralized training process for data and strong encryption processes, which complement international data protection standards. Blockchain-based systems also perform well, especially in complying with regulations (90%) and data integrity, using immutable records and distributed control. Compared to the Fuzzy NN model, which possesses moderate compliance (78%) and security (70%), it is deprived of advanced privacy-preserving methods. The IoT-Based System has the lowest rating in all categories, with a privacy rating of 65%, primarily because of edge device vulnerabilities and a lack of uniform security measures. The chart overall reflects the enhanced compliance potential of SafeLogFL as a better solution for secure and compliant cross-border logistics operations.

Fig. 8.

Privacy, security, and compliance comparison analysis.

Scalability and stakeholder collaboration

Scalability score (SS) is the system’s capacity to process growing volumes—greater data, devices, territories, or players—without compromising performance. It is obtained from Eq. 10.

10

where Response time (e.g., model training time or latency), Number of data sources or devices, and Number of stakeholders.

Stakeholder Collaboration Index (SCI) refers to how effectively multiple participants (e.g., customs agencies, carriers, logistics centers, data owners) can provide input, exchange information, and derive value from the system, trusting the system and keeping data confidential. This is obtained from Eq. 11.

11

where Interoperability score (0–1), ​ Agreement rate (e.g., model update acceptance), ​ Data transparency level, ​ Integration complexity, and Communication latency.

The Scalability Score (SS) and Stakeholder Collaboration Index (SCI) equations are shown to address assessment gaps in cross-border logistics risk warning systems that traditional federated learning (FL) measures cannot detect. Some standard metrics, such as accuracy, convergence, and communication cost, don’t demonstrate how well a system remains efficient as the number of users and data sources increases, or how effectively different stakeholders can collaborate to protect privacy. There are two types of equations: the SCI equation measures how well customs agencies, carriers, and logistics centers work together by looking at interoperability, agreement rate, transparency, integration complexity, and communication latency. The SS equation, on the other hand, measures the scalability of decentralized logistics environments by examining response time, the number of devices, and the number of stakeholders. When standard FL measures are insufficient in federated, privacy-preserving logistics networks, these domain-specific formulas, which are versions of distributed systems evaluation principles, ensure that scalability and teamwork can be measured.

Figure 9 (a) illustrates the Scalability and Stakeholder Collaboration of four cross-border logistics risk warning systems: SafeLogFL, Fuzzy NN, Blockchain-Based, and IoT-Based. From the left panel, SafeLogFL has the maximum scalability score of 85%, showing that it can handle more data and stakeholders better, followed by Blockchain-Based with 78%. As opposed to these, Fuzzy NN and IoT-Based systems display minimum scalability scores of 65% and 60%, respectively. Figure 9 (b) shows that SafeLogFL is also at the top in stakeholder cooperation (90%). This confirms its ability to facilitate effective collaboration between various parties due to its decentralized, privacy-preserving nature. Blockchain-Based ranks second (88%), with the advantage of transparency and audit ability, while Fuzzy NN and IoT-Based rank lower in this area (70% and 65%, respectively), reflecting information sharing and coordination difficulty. Overall, these findings portray SafeLogFL as the most collaborative and scalable system, thus optimal for dealing with cross-border logistics risks.

Fig. 9.

Comparison of scalability and stakeholder collaboration across cross-border logistics systems.

Computational efficiency

Computational Efficiency (CE) measures a system’s effectiveness in using resources like processing capacity, memory, and time with performance conservation. In cross-border warning systems for logistics, computational efficiency is critical because these systems would likely include large volumes of data, various stakeholders, and real-time decisions. An extremely efficient system performs the information processing task at high speed with very low computational overhead, which becomes inevitable in keeping federated learning models, blockchain, and IoT systems running, where computation may be a bottleneck. CE is a ratio of the performance output (such as prediction accuracy or model accuracy) to the resources consumed (such as time, energy, or computational load) and is shown in Eq. 12.

12

where refers to the system’s performance output, which could be model accuracy, prediction precision, or another relevant metric. the computational cost could include time (latency), memory usage, or power consumption.

Figure 10 compares the performance, computational expense, and scalability of four approaches: SafeLogFL, Fuzzy NN, Blockchain-Based, and IoT-Based. One set of bars represents each approach, and there are three bars per set for the measures: performance, computational expense, and scalability. The light salmon measure of performance shows how each approach performs, with taller bars indicating the best performance. Computational cost, light sea green, is how much resource every method consumes, with shorter bars preferred for optimization. Scalability, khaki, is how well each technique can scale with more data or tasks, with taller bars preferred to be more scalable. The figure facilitates direct comparison of methods, with SafeLogFL being the most optimized and Fuzzy NN being the most scalable. Blockchain-based and IoT-based approaches, being middle-range performance-wise, have mixed results regarding scalability and cost, with IoT-based being less efficient. This graph assists in ascertaining the strengths and compromises of each approach when it comes to the metrics, which can be used for decision-making according to the most significant factors of an application.

Fig. 10.

Computational efficiency analysis.

Conclusion and future work

The SafeLogFL framework envisioned addresses these issues in its stride with Federated Learning (FL) whereby various logistics stakeholders like customs bodies, shipping companies, and port terminals can collaboratively train prediction models without the sharing of sensitive information. With a Multi-Layer Perceptron (MLP) neural network integrated into a decentralized environment, SafeLogFL guarantees that each participant can train and process models on their respective local datasets. The locally trained models are then securely aggregated through the Federated Averaging (FedAvg) algorithm, maintaining data confidentiality without compromising the accuracy and generalizability of the model. The system proved highly accurate with a prediction rate of 91.3% in identifying risks like shipment delays, regulatory non-compliance, and disruptions. SafeLogFL is demonstrated to be scalable, GDPR-friendly, and responsive to real-time logistics settings. Its decentralized architecture fosters stakeholder confidence and enables a quick response to changing global trade trends. SafeLogFL thus establishes a standard for privacy-centric, intelligent logistics platforms capable of handling intricate international supply chains securely and effectively. The architecture presents a paradigm-shifting solution to managing logistics risk, striking a balance between predictive performance, data sovereignty, and global cooperation. Future effort will be directed towards improving SafeLogFL by adding real-time IoT data streams and investigating more sophisticated privacy-preserving methods such as secure multi-party computation for additional risk prediction accuracy and security improvement.

Author contributions

L.XW writing original draft preparation & methodology, L.XW investigation & writing review and editing.

Data availability

All data generated or analysed during this study are included in this article.

Declarations

Competing interests

The authors declare no competing interests.