Federated Learning-Based Framework: A New Paradigm Proposed for Supply Chain Risk Management ^†

Abstract

This paper proposes federated learning-based frameworks for supply chain risk management to address data-sharing constraints. To validate, centralized federated learning with horizontal data was applied for delivery delay prediction using datasets from two textile suppliers: supplier 1 has less data and is considered small, while supplier 2, with more data, represents a larger one. The prediction model is developed using an artificial neural network within the federated framework. The results show that federated learning benefits suppliers, especially the ones with limited data. Notably, federated learning outperforms centralized learning and local standalone learning. This highlights its potential to address privacy and facilitate collaboration.

1. Introduction

In the past decade, disruptions have profoundly affected industries worldwide, including global supply chains (SCs). As global trade grows and logistics networks become more complex, risks and uncertainties rise [1]. The SC disruption has become the most concerning topic in academia and industries, even after the COVID-19 pandemic. In SCs, disruptive threats can be divided into two categories: external risks, which originate from natural events, political conditions, competitors, and market fluctuations, and internal risks, which are tied to capacity, information systems, and internal operations [2]. Addressing these challenges adds complexity to managing risks effectively within SCs; however, it is vital to all participants.

Particularly, disruptions can cause delays. For suppliers, delivering products late compared to the designated date on the contract may result in contract violations and harm their reputation, especially if delays occur repeatedly. On the customer side, receiving products late can lead to sales losses or decreased satisfaction with the supplier. Considered low-impact but high-frequency events, delivery delays lasting up to two weeks may cause a sales reduction of as much as 10% within 3–4 weeks [3]. Especially in industries characterized by short product life cycles, high volatility, low predictability, and high-impulse purchasing, such as textiles and fashion, the need to address SC risks efficiently and promote SC resilience presents heightened challenges for SC risk management (SCRM) [4]. Due to the perishable nature of fashion products and the unpredictability of their demand, it is vital to maintain a responsive SC to swiftly deliver the products to the stores as needed [5]. To mitigate the risk of disruption, proactive strategies can help. Hence, anticipating potential delivery delays plays a critical role for SC stakeholders in improving production, inventory planning, and supporting what-if scenario execution [6].

SCRM is crucial for business continuity in the 4.0 Industry, where digital technologies have emerged as promising tools for enhancing SC management. As global SCs grow more complex, more advanced and collaborative approaches involving stakeholders are essential [7]. The traditional machine learning (ML) models require large, centralized datasets. However, the centralized ML approaches have proven inadequate in overcoming the challenges of privacy concerns and data isolation issues [8]. These limitations have led to the emergence of federated learning (FL) as a viable alternative. While FL has been widely applied in various domains, its application in SC management remains scarce [7,9,10].

Hence, this paper proposes FL frameworks aimed at enhancing SCRM, specifically addressing two key research questions: RQ1: How can FL enhance risk management in a SC? And RQ2: In which SC contexts can the FL method be applicable?

The remainder of the paper is structured as follows: Section 2 reviews the background and related works, exploring SCRM within the Industry 4.0 context and its challenges. Section 3 presents the proposed FL frameworks. Section 4 describes an experiment predicting delivery risks in the textile industry using centralized federated learning. Section 5 concludes the paper, summarizing the findings and limitations and suggesting future research directions.

2. Background and Related Works

Industry 4.0 highlights the challenges posed by increasing complexity, data privacy concerns, and the need for advanced technologies to enhance resilience. The field of SCRM focuses on risk management techniques designed for identifying, assessing, mitigating, and monitoring risks within the SC [11]. These risks can stem from microscopic and macroscopic sources. The goal of SCRM is to build a resilient SC that can adapt to changes and disruptions while minimizing adverse impacts on business operations. Since the COVID-19 outbreak, SC disruption has gained considerable attention, making its prediction important for SC stability. In the context of Industry 4.0, digital technology has increasingly emerged as a powerful tool for enhancing the effectiveness and efficiency of SCRM. The ability of artificial intelligence and ML to process massive datasets, uncover patterns, and predict future outcomes with exceptional accuracy makes them a compelling option [12]. Moreover, using digital technologies helps companies manage risks in their SC better, which then significantly affects corporate circular economy capability [13]. ML enables organizations to identify high-risk suppliers and predict delays, allowing them to take proactive strategies to mitigate the impact of disruptions, enhance resilience, and improve SC efficiency [14]. Within SCRM, data sharing aims to enable learning from past disruption events and supports the assessment of risks shared by all parties [4,7]. Unfortunately, most of the existing studies on information sharing concentrate on the benefits of shared data, assuming that institutions are willing to share, while neglecting the influence of their willingness to share [15]. Furthermore, conventional AI and ML models require centralized data, posing risks of organizational exposure and opportunistic use [7].

In an Industry 4.0 context, advancements in modern information and communication technology have greatly facilitated the generation, collection, and processing of personal data for various purposes, raising significant privacy concerns [16]. Two major challenges in data science development are data governance and data silos, where data governance holds the greatest importance [17]. Data governance laws, such as the General Data Protection Regulation (GDPR), which has been effective since 25 May 2018, grant users full ownership of their data, prohibiting its use without their consent. Similarly, the California Consumer Privacy Act (CCPA) has set high privacy standards, influencing global data protection and reshaping how organizations manage personal data. As a result, stakeholders in SCs may encounter legal challenges related to data sharing when collaborating in SCRM. The other major challenge, that of data silos, where data are isolated within systems, departments, or organizations, hinders industry development by limiting access to larger datasets that are essential for improving ML model performance. These challenges are directly related to data sharing and collaboration along the SC. Collaboration among SC members enhances resilience to disruptions through shared information, aligned targets, integrated decisions, efficient communication, resource sharing, creativity, and incentive alignments, ensuring a unified and adaptive response to disruptions [18]. Data sharing has long been recognized as a critical mechanism for mitigating undesirable effects such as the Bullwhip effect and shortage gaming, but the field of SC management has long struggled with this issue [10], as concerns about confidentiality remain a significant barrier [7].

FL is a promising new approach that aids in developing privacy-preserving systems and secure distributed learning models. Since 2022, FL has gained popularity in SCRM [6]. Ref. [10] proposed centralized federated learning (CFL) for SC risk prediction, ensuring data privacy from three buyers in the UK maritime sector. Ref. [9] suggested a CFL framework for order-level risk prediction in aerospace, while [19] applied CFL for credit card fraud detection. Ref. [10] also developed an adaptive FL approach to predict delivery delays in e-commerce SCs. However, the application of FL in SC management is still scarce [7,9].

3. Federated Learning and Taxonomy

FL, postulated by Google in 2016, is a learning technique that enables collaborative learning from decentralized data while maintaining data privacy by avoiding central storage [20]. The process of FL has three basic steps as follows: (i) each entity locally trains a model on its own dataset; (ii) each entity sends its updated model to a designated aggregation point; and (iii) the aggregation point combines the received model updates to form the global model. Today, FL is the leading framework for collaboratively training ML models while maintaining privacy constraints [21]. In SCRM, FL can offer several advantages, such as (i) decentralized training without sharing raw data, reducing data storage complexities; (ii) enhanced privacy; (iii) cost reductions; (iv) lower latency; (v) reduced power consumption; and (vi) collaborative organizational action [6]. The following sessions explain the main FL types.

3.1. Centralized Federated Learning

CFL is the conventional form of FL, where a central server aggregates local models from all entities to form a global model. The framework of CFL is illustrated in Figure 1, where $E=\left\{E_1,E_2,\dots ,E_k,\dots ,E_n\right\}$ represents the data owners, each with their respective dataset $D=\left\{D_1,D_2,\dots ,D_k,\dots ,D_n\right\}$. The CFL process begins with (i) the initialization of the global model weights ($w_t^{glo}$) in the central server, and sends the global model to each entity. Next, (ii) each entity $E_k$ trains its model locally without exposing its data $D_k$ to others, followed by (iii) sending the updates ($w_{1,t}^{lo}$, $w_{2,t}^{lo}$, …, $w_{n,t}^{lo}$) to a central server. (iv) The server aggregates these updates into a global model and (v) sends it back to the entities. Finally, (vi) each entity incorporates the global update into its local model. During the FL process, FedAvg (Federated Averaging) is known as a popular and effective aggregation algorithm that combines local models through parameter-wise averaging [22]. This aggregation method uses Unweighted FedAvg, given by $W^{glo}={\displaystyle \frac{1}{N}}{\sum }_{i=1}^N{W}_i$ and Weighted FedAvg, expressed as $W^{glo}={\displaystyle \frac{{\sum }_{i=1}^N{n_{i}W}_i}{{\sum }_{i=1}^N{n}_i}}$, where N represents the total number of entities, and $n_i$ denotes the number of data points used by entity i. Unweighted FedAvg treats all entities equally during model aggregation, regardless of the size of their local datasets. In contrast, Weighted FedAvg adjusts each entity’s contribution based on the size of their dataset, giving more influence to the entities with larger datasets.

FL is still emerging, with most applications using the widely adopted CFL architecture [22]. CFL is easier to implement in the SC and works well in the context that SC partners trust a central server for aggregation. CFL has attracted considerable attention from both academia and industry [23].

3.2. Decentralized Federated Learning

While CFL has been widely used, it has limitations, such as communication bottlenecks and central server reliance. Decentralized FL (DFL), introduced in 2018, tackles these challenges by facilitating decentralized model aggregation and minimizing reliance on a central server [22]. Generally, DFL operates as follows: (i) entities begin by initializing their local models independently; (ii) they then train these models locally using their own datasets; (iii) next, entities directly exchange model parameters with other ones via peer-to-peer (P2P) communication; and (iv) finally, they refine their local models by aggregating the received parameters from their peers. DFL becomes more feasible in the context that stakeholders prefer direct P2P collaboration rather than relying on a central server.

3.3. Semi-Centralized Federated Learning

As CFL often imposes significant computational demands on the server, while DFL incurs substantial P2P communication costs [24], semi-centralized FL (SCFL) was introduced as a hybrid model, aiming to balance the benefits and drawbacks of both centralized and decentralized approaches, optimizing collaboration while maintaining data privacy. In SCFL, entities locally train models on their own datasets and update their models to sub-servers (agents) using the CFL approach, while the global model is aggregated through DFL among sub-servers. The SCFL methods can be categorized into (i) clustered FL and (ii) hierarchical FL [24]. In clustered FL, data owners are grouped into clusters to obtain the global model, while in hierarchical FL, local models are trained by data nodes, aggregated by intermediate nodes, and then further aggregated to form the final federated model.

3.4. Semi-Decentralized Federated Learning

In contrast to CFL, DFL, and SCFL, a hybrid architecture that combines aspects of both CFL and DFL, where local models use DFL—but the role of the aggregator for forming the global model rotates randomly among participants at each round of the learning process to eliminate potential single points of failure and improve overall system resilience—is called semi-decentralized FL (SDFL). SDFL is a cost-effective solution for larger networks, providing comparable performance to DFL while requiring fewer resources [22]. Both SCFL and SDFL may be ideal for a complex SC that involves multiple regional hubs (across different geographical regions) or sectors (encompassing different parts of the SC, such as sourcing, manufacturing, logistics, etc.).

3.5. Data in FL

FL can be categorized into three types based on data distribution: horizontal FL (HFL), vertical FL (VFL), and federated transfer learning (FTL) [25]. The FL comprehensive training dataset is represented as (X, Y, I), where X is the feature space, Y is the label space, and I is the sample identity space. Accordingly, HFL applies to datasets where ${\mathrm{X}}_{\mathrm{i}}={\mathrm{X}}_{\mathrm{j}},{\mathrm{Y}}_{\mathrm{i}}={\mathrm{Y}}_{\mathrm{j}},{\mathrm{I}}_{\mathrm{i}}\ne {\mathrm{I}}_{\mathrm{j}}\forall {\mathrm{D}}_{\mathrm{i}},{\mathrm{D}}_{\mathrm{j}},\mathrm{i}\ne \mathrm{j}$; VFL applies to datasets where ${\mathrm{X}}_{\mathrm{i}}{\ne \mathrm{X}}_{\mathrm{j}},{\mathrm{Y}}_{\mathrm{i}}\ne {\mathrm{Y}}_{\mathrm{j}},{\mathrm{I}}_{\mathrm{i}}={\mathrm{I}}_{\mathrm{j}}\forall {\mathrm{D}}_{\mathrm{i}},{\mathrm{D}}_{\mathrm{j}},\mathrm{i}\ne \mathrm{j}$; and FTL applies when ${\mathrm{X}}_{\mathrm{i}}\ne {\mathrm{X}}_{\mathrm{j}},{\mathrm{Y}}_{\mathrm{i}}\ne {\mathrm{Y}}_{\mathrm{j}},{\mathrm{I}}_{\mathrm{i}}\ne {\mathrm{I}}_{\mathrm{j}}\forall {\mathrm{D}}_{\mathrm{i}},{\mathrm{D}}_{\mathrm{j}},\mathrm{i}\ne \mathrm{j}$.

4. An Experiment: Predicting the Delivery Risk in the Textile Supply Chain

This study proposes FL-based frameworks for SCRM, addressing data privacy, decentralized data, and model training challenges. To validate FL’s application, we conducted an experiment using CFL to predict delivery delays for two textile suppliers.

4.1. Data Description

For the delivery delay prediction, the key features derived from [7,10] are used, as shown in

Table 1. Features selected and description.

Feature	Description	Data Type
Scheduled_Days	Days of scheduled delivery of the purchased product	Numeric
Actual_Days	Actual shipping days of the purchased product	Numeric
Delivery_Risk	∆DR represents the difference between “Actual_Days” and “Scheduled_Days” with a value of 1 for “Late” and 0 for “On-time”.	Numeric
Order_Date	The date on which the order is made	Date/Time
Shipping_Date	The exact date and time of shipment	Date/Time
Order_Status	Status of an order delivery includes complete, pending, closed, pending payment, canceled, processing, suspected fraud, on hold, and payment review	Categorical
Shipping_Mode	The following shipping modes are presented: Standard class, first class, second class, and same day	Categorical

. The datasets from Mendeley Data [26] include 12,487 observations for supplier 1 (S1) and 22,246 for supplier 2 (S2). The key features include “Scheduled_Days”, “Actual_Days”, “Order_Date”, “Shipping_Date”, “Order_Status”, and “Shipping_Mode”. Delivery risk (∆DR) is defined as the difference between “Actual_Days” and “Scheduled_Days”, and if ∆DR is positive, it means “Late”; otherwise, it means “On-time”. Label encoding is applied to categorical features such as “Order_Status” and “Shipping_Mode”. We normalized the data using min-max normalization, defined as: $x_i^{normalized}={\displaystyle \frac{x_i-\mathrm{m}\mathrm{i}\mathrm{n}\left(x\right)}{\max \left(x\right)-\mathrm{m}\mathrm{i}\mathrm{n}\left(x\right)}}$.

Table 2. Summarization of dataset homogeneity testing.

	Kolmogorov–Smirnov (K-S) Test ${\mathbf{H}}_0$: The Distributions of the Two Datasets Are Identical				Chi-Square Test ${\mathbf{H}}_0$: The Distributions of the Two Datasets Are Identical
Features	Scheduled_Days	Delivery_Risk	Order_Date	Shipping_Date	Order_Status	Shipping_Mode
Test Statistic	0.0031	0.0001	0.0217	0.0216	9.5394	1.7707
p-value	0.9999	1.0	0.0011	0.0012	0.2988	0.6213

presents the homogeneity test results for the two datasets used in prediction. The Kolmogorov–Smirnov (K-S) test is applied to the numeric features, while the Chi-Square test is used for the categorical features. The results indicate that “Scheduled_Days”, “Delivery_Risk”, “Order_Status”, and “Shipping_Mode” have similar distributions (p-value > 5%), while “Order_Date” and “Shipping_Date” differ significantly (p-value < 5%). In FL, data heterogeneity often leads to client drift, harming model performance [27].

Figure 2 illustrates feature importance using XGBoost, showing that “Shipping_Mode” is the strongest predictor of “Delivery_Risk”, followed by “Scheduled_Days”, aligning with the findings from [10]. “Order_Status” and “Order_Date” follow in importance, while “Shipping_Date” has the least impact.

4.2. Model Setting and Performance Metrics

To train the model in FL, Artificial Neural Networks (ANNs), inspired by the human brain’s structure and function, are used. Today, the ANN has become more popular because of its ability to learn from data and deliver accurate predictions [28]. As shown in

Table 3. Parameters and architecture of the ANN model in the CFL framework.

Layer	Units	Activation Function	Regularization	Remarks
Dense Layer 1	32	ReLU	L2 regularization (0.0001)	First hidden layer
Dense Layer 2	16	ReLU	L2 regularization (0.0001)	Second hidden layer
Dense Layer 3	8	ReLU	L2 regularization (0.0001)	Third hidden layer
Output Layer	1	Sigmoid	-	Binary classification (0 or 1)

, the ANN model is designed for binary classification tasks and is built using the Keras Sequential API within a FL framework to collectively predict delivery risks. It comprises an input layer (32 neurons), two hidden layers (16 and 8 neurons with ReLU activation), and a Sigmoid output layer with one neuron for binary classification. The model trains for 100 epochs with a batch size of 32, learning risk patterns while preserving data privacy. The model is compiled with the Adam optimizer, a learning rate of 0.001, and binary cross-entropy as the loss function, while accuracy is used as the evaluation metric. To prevent overfitting, L2 regularization is applied to each layer, penalizing large weights [29].

Table 4. CFL algorithm.

1:	Initialize the global model weights $W_0^{gl}$ in the main server
2:	$\mathrm{Distribute}{W}_0^{gl}$ to all entities $k\in \left\{\mathrm{1,2},\dots ,n\right\}:$${W_{k,0}^{lo}=W}_0^{gl}$
3:	For each round $t\in \left\{\mathrm{1,2},\dots ,T\right\}:$
4:	Each entity receives the global model: ${W_{k,t}^{lo}=W}_t^{gl}$
5:	Each entity k locally trains ANN model: $W_{t+1}^{lo} ⃪ LocalUpdate$$(D_k,$$W_{k,t}^{lo})$.
6:	The entity k sends the updated local model $W_{k,t+1}^{lo}$ to the main server.
7:	Aggregate local weights: $W_{t+1}^{gl}\leftarrow f\left(\sum _{k=1}^n{W}_{k,t+1}^{lo}\right)$.
8:	$\mathrm{Send}{W}_{t+1}^{gl}$ to all entities for the next round.
9:	Repeat until t = T.

shows the algorithm for the CFL framework for the prediction. Additionally, to address the issue of imbalanced training data, the FL system incorporates the Synthetic Minority Over-sampling Technique (SMOTE), as emphasized in [30].

We evaluated the model performance using four metrics—accuracy, precision, recall, and the F1-score—from the confusion matrix (

Table 5. Structure of the confusion matrix.

		Predicted label
		Positive	Negative
True label	Positive	True Positive (TP)	False Negative (FN)
	Negative	False Positive (FP)	True Negative (TN)

), defined as follows:$\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}={\displaystyle \frac{\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}+\mathrm{F}\mathrm{N}+\mathrm{T}\mathrm{N}}}$; $\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}}}$; $\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}}$ and $\mathrm{F}-\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}={\displaystyle \frac{2\ast \mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\ast \mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}{\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}+\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}}$.

4.3. Results and Discussion

Table 6. Accuracy, precision, recall, and F1-score of LSL, CL, and CFL.

Metrics	CL	LSL		CFL.Weighted		CFL.Unweighted
		S1	S2	S1	S2	S1	S2
Accuracy	0.7846	0.7700	0.7810	0.7899	0.7855	0.7913	0.7868
Precision	0.7851	0.7720	0.7818	0.7901	0.7860	0.7916	0.7872
Recall	0.7846	0.7700	0.7810	0.7899	0.7855	0.7913	0.7868
F1-score	0.7846	0.7696	0.7808	0.7898	0.7854	0.7913	0.7867

presents the model’s performance metrics, highlighting that CFL outperforms other scenarios, including centralized learning (CL) and local standalone learning (LSL). Notably, CFL using unweighted federated averaging for global model aggregation achieves the best performance across all metrics. Compared to LSL, CL performs better; nonetheless, this scenario is not practical in real-world SCs. Additionally, S1, with less data, exhibits the lowest performance across all metrics, while S2, with more data, demonstrates better overall performance. Figure 3 illustrates the performance differences between LSL and CL, showing that the differences (LSL-CL) for S2 are consistently smaller than those for S1. This suggests that enterprises with fewer data, typically small and medium-sized ones (SMEs), tend to achieve more limited model training outcomes compared to larger enterprises, as emphasized by [7]. Importantly, the results reveal that S1 can significantly improve its prediction model, while S2 also experiences a performance increase when they participate in FL together. Compared to LSL for S1, CFL can, on average, improve accuracy by 0.0206, precision by 0.0189, recall by 0.0209, and the F1-score by 0.0210. Our study indicates that SMEs significantly benefit from FL, as concluded by [7,9]. Remarkably, however, this study shows that CFL outperforms CL, with the performance differences between CFL and CL being consistently positive (Figure 3).

Figure 4, Figure 5, Figure 6, Figure 7, Figure 8, Figure 9 and Figure 10 present the confusion matrices for model classification, while

Table 7. Precision, recall, and the F1-score from LSL, CL, and CFL.

Metrics	CL		LSL				CFL.Weighted				CFL.Unweighted
			S1		S2		S1		S2		S1		S2
	On-Time	Late	On-Time	Late	On-Time	Late	On-Time	Late	On-Time	Late	On-Time	Late	On-Time	Late
Precision	0.7730	0.7972	0.7488	0.7952	0.7663	0.7974	0.7820	0.7982	0.7747	0.7973	0.7830	0.8002	0.7765	0.7978
Recall	0.8059	0.7633	0.8126	0.7274	0.8087	0.7533	0.8038	0.7759	0.8054	0.7657	0.8060	0.7766	0.8054	0.7682
F1-score	0.7891	0.7799	0.7794	0.7598	0.7869	0.7747	0.7928	0.7869	0.7897	0.7811	0.7944	0.7882	0.7907	0.7827

displays the classification report, including precision, recall, and the F1-score for each class (1: “Late”, 0: “On-time”) across two suppliers and three scenarios. The results show that CFL (both Weighted FedAvg and Unweighted FedAvg for global model updates) generally outperforms CL and LSL across the performance metrics and suppliers. This aligns with the findings of [31], which highlighted that FL delivers superior performance compared to CL. In this study, CFL provides a more balanced and consistent performance, highlighting its effectiveness in improving model accuracy. The F1-score for the “Late” class of S1 increases from 0.7598 with standalone training to 0.7869 and 0.7882 when participating in FL. Similarly, S2’s F1-score for the “Late” class rises from 0.7747 with its own data to 0.7811 and 0.7827 with FL participation. Interestingly, the performance difference between S1 and S2 is minimal for CFL, compared to LSL, highlighting how FL helps reduce data inequality and enhance prediction model performance while ensuring data privacy.

The delivery risk prediction model for suppliers S1 and S2 demonstrates steady improvement. Both training and validation accuracy gradually increase, stabilizing between 77.5% and 80% after 100 epochs. Similarly, the training and validation loss consistently decrease, reaching a plateau between 0.40 and 0.45 after approximately 60 epochs. The convergence and alignment of the accuracy and loss curves indicate the model’s reliability, generalization capability, and robustness.

5. Conclusions

This paper proposes FL-based frameworks for SCRM, addressing privacy concerns and data-sharing constraints in Industry 4.0 SCs. We validated the framework using CFL with horizontal data to predict delivery delays in the textile industry, leveraging datasets from two suppliers of varying sizes. The results from the ANN model within the CFL framework show that FL provides advantages for two suppliers, especially for suppliers with limited data, reducing data inequality. Notably, our study reveals that CFL outperforms CL and LSL, demonstrating its feasibility as a privacy-preserving alternative compared to traditional learning methods. This confirms FL is a promising approach for enhancing risk management through knowledge sharing, collaboration, and fostering SC resilience.

Limitations and Further Study

The choice of ML algorithm in FL affects prediction performance, and future research should explore various algorithms. While FedAvg is simple, it may not suit all participants with differing data distributions, highlighting the need for personalized FL. Moreover, in Industry 4.0, integrating Blockchain, Internet of Things (IoT), and FL provides a new avenue for enhancing SC collaboration, thus boosting efficiency, transparency, and resilience. Additionally, understanding barriers and enablers is crucial for successful FL adoption and requires further exploration.