|August October 2023

In principle, FL can be an extremely useful technique to address critical issues of industrial IoT (IIoT) applications. As such, it matches perfectly [[ToloMEO Embedded Assistant|https://tolomeo.io DAVE Embedded Systems' IIoT platform, ToloMEO]]. This Technical Note (TN) illustrates how DAVE Embedded Systems explored, tested, and characterized some of the most promising open-source FL frameworks available to date. One of these frameworks might equip ToloMEO-compliant products in the future allowing our customers to implement federated learning systems easily. From the point of view of machine learning, therefore, we investigated if typical embedded architectures used today for industrial applications are suited for acting not only as inference platforms — we already dealt with this issue [[ML-TN-001 - AI at the edge: comparison of different embedded platforms - Part 1|here]] — but as training platforms as well.

* '''ML frameworks flexibility''': The adaptability of the framework to manage different ML frameworks.* '''Licensing''': It is mandatory that the framework has an open-source, permissive license to cope with the typical requirements of real-world use cases.* '''Repository rating and releases''': Rating in a repository is important for a FL framework as it indicates a high level of community interest and support, potentially leading to more contributions and improvements. Meanwhile, the first and latest releases indicate respectively the maturity and the support of the framework and whether it is released or still in a beta version.* '''Documentation and tutorials''': The provided documentation with related tutorials has to be complete and well-made.* '''Readiness for commercial usage''': The readiness of the framework to be developed in a real-world scenario. In order to establish the readiness, it was checked the version of the framework and the license.

The choice of a suitable license is of paramount importance for any FL framework. A well-crafted license provides a legal foundation that governs the usage, distribution, and modification of the framework’s source code and associated components. A permissive license, like the MIT License or Apache License, allows users to use, modify, and distribute the framework with relatively few restrictions. This encourages widespread adoption, fosters innovation, and facilitates contributions from a broader community of developers and researchers. The permissiveness of these licenses empowers users to incorporate the framework into their projects, even if hey have proprietary components. On the other hand, copyleft licenses, like the GNU GPL, require derived works to be distributed under the same terms, ensuring that any modifications or extensions to the framework remain open-source. While this may be more restrictive, it encourages a collaborative ecosystem where improvements are shared back with the community. A clear and well-defined license also provides legal protection to both developers and users, helping to mitigate potential legal risks and disputes. It ensures that contributors have granted appropriate rights to their work and helps maintain a healthy and sustainable development environment. Most of the frameworks previously described are under the Apache-2.0 license except one: IBMFL. In fact, it is under an unspecified license that makes the framework not suitable for commercial use. For that reason, IBMFL was discarded from the comparison too.

* '''Visibility''': Repositories with good ratings are likely to appear higher in platform's search results, making it easier for others to discover and use the project.* '''Credibility''': High-rating repositories are often perceived as more trust-worthy and reliable, as they are vetted and endorsed by a larger user base.* '''Contributions''': Popular repositories tend to attract more contributions from developers, leading to a more active and vibrant community around the project.* '''Feedback''': Projects with good ratings are more likely to receive feedback, bug reports, and feature requests, helping the developers improve the software.* '''Maintenance''': Higher ratings can also stimulate the maintainers to keep the project updated and actively supported. Other important, rating-related aspects are the first and latest releases. Thanks to the latter, it is possible respectively to see the maturity of the framework and also how often it is updated, and thus the support behind it. Obviously, a framework that was born earlier than others is much more likely to have better ratings. Having this in mind, at the time of writing this thesis, the ranking in terms of received stars correlated with the first release for each framework is as follows:** '''PySyft''': 8.9k stars / Jan 19, 2020** '''FATE''': 5.1k stars / Feb 18, 2019** '''FedML''': 3.1k stars / Apr 30, 2022** '''Flower''': 2.8k stars / Nov 11, 2020 ** '''TFF''': 2.1k stars / Feb 20, 2019 ** '''OpenFL''': 567 stars / Feb 1, 2021 ** '''IBMFL''': 438 stars / Aug 28, 2020 ** '''NVFlare ''': 413 stars / Nov 23, 2021

* '''Accessibility and Ease of Use''': Comprehensive documentation allows users to understand the framework’s functionalities, APIs, and usage quickly. It enables developers, researchers, and practitioners to get started with the framework efficiently, reducing the learning curve.* '''Accelerated Development''': Well-structured tutorials and examples demonstrate how to use the framework to build practical FL systems. They provide step-by-step guidance on setting up experiments, running code, and interpreting results. This expedites the development process and encourages experimentation with different configurations.* '''Error Prevention''': Clear documentation and good examples help users avoid common mistakes and errors during implementation. It provides troubleshooting tips and addresses frequently asked questions, reducing frustration and increasing user satisfaction.* '''Reliability and Robustness''': A well-documented framework indicates that developers have invested time in organizing their code and explaining its functionalities. This attention to detail suggests a more reliable and stable framework.* '''Maintenance''': Higher stars can also stimulate the maintainers to keep the project updated and actively supported.

At the beginning of this section, a total of eight frameworks were considered. Each framework was assessed based on various aspects and after an in-depth analysis, six frameworks were deemed unsuitable due to some requisites not being met. The requirements that were considered are summarized in the following table:

These two remaining frameworks are then: '''Flower ''' and '''NVFlare'''. They demonstrated the potential to address the research objectives effectively and were well-aligned with the specific requirements of the FL project. Later, these two selected frameworks will be rigorously compared, examining their capabilities in handling diverse ML models, supporting various communication protocols, and accommodating heterogeneous client configurations. The comparison will delve into the frameworks’ performance, ease of integration, and potential for real-world deployment. By focusing on these two frameworks, this research aims to provide a detailed evaluation that can serve as a valuable resource for practitioners and researchers seeking to implement FL in a variety of scenarios. The selected frameworks will undergo comprehensive testing and analysis, enabling the subsequent sections to present an informed and insightful comparison, shedding light on their respective strengths and limitations.

cars, birds, etc.), the Cross-Entropy loss compares the predicted class probabilities with the true one-hot encoded labels for each input sample. It applies the logarithm to the probabilities and then sums up the negative log likelihoods across all classes. The objective is to minimize this loss function during the training process, which effectively encourages the model to assign high probabilities to the correct class labels and low probabilities to the incorrect ones. One of the reasons why Cross-Entropy Loss is considered suitable for CIFAR-10 and classification tasks, in general, is its ability to handle multi-class scenarios efficiently. By transforming the model’s output into probabilities through the softmax activation, it inherently captures the relationships between different classes, allowing for a more expressive representation of class likelihoods.

* '''Loss''': The the loss function quantifies the dissimilarity between the predicted output of the model and the actual ground truth labels in the training data. It provides a measure of how well the model is performing during training. The goal is to minimize the loss function, as a lower loss indicates that the model is better aligned with the training data.* '''Accuracy''': Accuracy accuracy is a fundamental metric used to assess the model’s overall performance. It represents the proportion of correctly predicted samples to the total number of samples in the dataset. A higher accuracy indicates that the model is making accurate predictions, while a lower accuracy suggests that the model might need further improvements. Calculating the accuracy of individual clients in a FL classification problem is important to assess the performance of each client’s local model. This helps in understanding how well each client is adapting to its local data distribution and making accurate predictions.* '''F1-score''': The the F1-score is a metric that combines both precision and recall to provide a balanced evaluation of the model’s performance, especially when dealing with imbalanced datasets. Precision measures the ratio of correctly predicted positive samples to all predicted positive samples, while recall measures the ratio of correctly predicted positive samples to all actual positive samples. The F1-score is the harmonic mean of precision and recall, providing a single metric that considers both aspects.

 {| class="wikitable" style="margin: 0 auto;"

* '''Global evaluation''': accuracy and F1-score at the end of each round of FL were tested.* '''Local evaluation''': accuracy and F1-score at the end of each round of FL were tested.* '''Training evaluation''': it was computed loss, accuracy, and F1-score.

* '''Limited hardware resources''': Embedded devices often have limited hardware resources, such as CPU, memory and computing power. This restriction can affect the performance of FL, especially if models are complex or operations require many resources.* '''Hardware variations''': Embedded devices may have hardware variations between them, even if they belong to the same class. These hardware differences may lead to different behaviors in FL models, requiring more robustness in adapting to different devices.* '''Variations in workload''': Embedded device applications may have very different workloads from those simulated in a virtual environment. These variations may lead to different requirements for FL.

{| class="wikitable" style="margin: 0 auto;"

To the end of performing more advanced testing, the same problem described in chapter "Flower vs NVFlare: an in-depth comparison" was leveraged. However, the test bed was tuned in order to increase the complexity of the use case as detailed in the rest of this chapter. The design settings of this advanced FL system remain consistent with those utilised utilized in the previous comparison between NVFlare and Flower, referred to  as the "Local Environment" in section 4.1.1, unless some changes. In this sce- narioscenario, the same desktop machine was utilisedutilized, equipped with an NVidia RTX 3080 Ti GPU. The ML framework, Pytorch, remained consistent, as did the Data Preprocessing involving Dataset selection, Dataset splitting, and Data augmentation. However, a significant alteration change was introduced between the initial two design components, elaborated upon in subsection 4.1.4, referred to as regarding "Data heterogeneity". The Model configuration and client-side set- tings settings also remained unchanged. Minor adjustments were made to the met- rics metrics taken into consideration, focusing exclusively on two: local training loss and server validation accuracy. On the server side, the configuration under- went underwent modifications. While maintaining a count of four clients, the number of communication rounds was elevated to 20 in this particular scenario. == FL Algorithms and Centralised Simulation ==One of the two main changes made to the system design was to simulate a centralised training baseline and to consider two other algorithms in addition to FedAvg. The centralised training was conducted using a single client for 20 local epochs, aiming to simulate a ML environment. This approach served as a reference point for comparison against various instances of FL.

== FL algorithms and centralized simulation ==One of the two main changes made to the system design was to simulate a centralized training baseline and to consider two other algorithms in addition to FedAvg. The centralized training was conducted using a single client for 20 local epochs, aiming to simulate a ML environment. This approach served as a reference point for comparison against various instances of FL. The other two FL algorithms employed in this study are Federated Op-Optimisation (FedProx) and Stochastic Controlled Averaging for FL (Scaffold). Starting with FedProx, this algorithm extends the conventional FedAvg method by introducing a proximal term. The proximal term adds a regularization factor to the optimization process, enhancing the convergence rate and stability of the model across participating clients. FedProx achieves this by optimizing the global model using both local updates and a global proximal term, which balances the contributions of individual clients while preventing divergence.

timisation (FedProx) [37] and Stochastic Controlled Averaging for FL (Scaf- fold) [52]. Starting with FedProxMoving on to Scaffold, this algorithm extends approach focuses on refining the conventional FedAvg method aggregation step of FL. It introduces controlled averaging by introducing employing the variance of model updates as a proximal termcontrol signal. The proximal term adds a regulari- sation factor This allows Scaffold to dynamically adjust the optimisation processaggregation weight of each client’s update based on their historical per-ormance. By doing so, enhancing the convergence rate and stability algorithm mitigates the effects of the model across participating clients. FedProx achieves this by optimising the global model using both local noisy updates and a global proximal term, which balances improving the contributions overall convergence of individual clients while preventing divergence. This can be better seen in the Listing 5FL process.1:

In this advanced project, an additional feature was incorporated involvingthe integration of classes aimed at performing dataset splitting among the designated clients, which, in this instance, were four in number. In addition to dividing the dataset into four subsets, the possibility of choosing the level of heterogeneity of the data was added by applying the Dirichlet sampling strategy. Thus, it was possible to dynamically adjust the degree of data heterogeneity for each client bringing higher. This functionality made it possible to simultaneously customize the level of data heterogeneity across all clients. In the context of FL, this data heterogeneity can be defined as follows:* '''Low Data Heterogeneity''': Low heterogeneity means that the data across different clients is quite similar or homogeneous. There is little variation among the data held by different clients. This leads to nearly balanced classes among clients, that is classes with a similar number of samples in each class.* '''High Data Heterogeneity''': High heterogeneity means that there is significant diversity in the data across different clients or nodes. This means that every subset assigned to each client contains unbalanced classes, i.e. some classes may be over-represented in some customers, while others may be under-represented.In order to have a clear comparison within the experiments, the upper and lower extremes of the α factor affecting heterogeneity were considered, i.e. 0.1 and 1.0.

the integration == Results analysis ==A series of classes aimed at performing dataset splitting among seven experiments were conducted. The first experiment involved a centralised simulation, while the designated clientsremaining six experiments focused on testing three different algorithms: FedAvg, which, in this instance, were four in numberFedProx and Scaffold. In addition to dividing the dataset into four subsetsSpecifically, each algorithm was tested twice: the possibility of choosing first time with α = 0.1 and the level of heterogeneity of the data [48] was added by applying the Dirichletsecond time with α = 1.0.

sampling strategy [36]The following figure represents the local <code>training_loss</code> obtained running the quoted experiments. As can be seen, the loss of the centralized simulation isn’t good enough to keep up with the other experiments that reach a bit lower loss values. This shows the effectiveness of the FL algorithms compared to a classical ML approach. It can also be noticed the same behavior obtained in section with the previous test bed, where at the beginning of each round the loss go instantly higher compared to the last epoch of the previous round to get lower with later epochs. Another important thing to note is how experiments with an alpha value of 0.1 perform worse than their counterpart evaluated from an alpha value of 1.0.

== Results analysis ==A series This factor becomes even more evident when observing the following chart, which illustrates the server <code>validation_accuracy</code>. This is due to the fact that there are more unbalanced classes within each client’s dataset and this leads models trained on classes with less data to have difficulty generalizing correctly. Models are more inclined to predict dominant classes, reducing accuracy on less represented classes. The poor representation of seven experiments were conductedsome classes makes it difficult for models to learn from them, leading to lower overall accuracy. The first experiment involved

a centralised simulation, while the remaining six experiments focused on test[[File:NVFlare-server-validation-accuracy.png|center|thumb|732x732px|NVFlare: server validation accuracy.]]

ing three different Analyzing the individual algorithms: , it can be seen a very similar behavior between FedAvgand FedProx, which have very similar results in terms of both local <code>training_loss</code> and server <code>validation_accuracy</code>. This is mainly due to the fact that they are very similar to each other minus a proximity term mu, in the case of FedProx , which improves the convergence ratio. The Scaffold algorithm, on the other hand, has a totally different implementation from its predecessors, which allows to dynamic adjustment of the aggregation weight of each client’s update based on their historical performance, and Scaffoldthus achieves better performance, especially when using unbalanced classes (α = 0.1). This can easily be seen in the server <code>validation_accuracy</code> graph. Specifically,

each algorithm was The successful execution of this more complex use case on NVFlare, involving multiple tested twice: algorithms and diverse data heterogeneity, further underscores the first time with α = 0framework’s robust capabilities and suitability for a wide range of scenarios.1 and This result confirms the second time versatility of NVFlare as a FL framework, making it a reliable choice for real-world scenarios dealing with α = 1.0heterogeneous and complex data.

TBD One The analysis detailed in previous chapter allows to say that NVFlare is a viable framework for building real-world Federated Learning systems that make use of Linux-powered embedded platforms as clients. Nevertheless, it is worth remembering that one important issue that was purposely not addressed , yet is . For the sake of simplicity, this work did not consider the problem of labeling of new samples. In other words, it was implicitly assumed that new samples collected by the clients are somehow labelled prior to being used for training. This is a strong assumption because implies . In reality, system architects can not overlook this fundamental issue when designing FL-based solutions. This is especially true in industrial environments where clients are often unattended devices that labeling of new samples issue* operate standalone. To date, this topic hes not been investigated thoroughly yet. For instance, [https://bdtechtalks.com/2021/08/09/what-is-federated-learning/* this article] mentions it without providing practical solutions, however. Generally speaking, it seems that for industrial applications the use of unsupervised learning?**techniques could be a promising approach (see for example [https://arxiv.org/pdf/1805.03911.pdfthis paper]). In any case, the problem of labeling new samples will have to be addressed in future works to make Federated Learning truly available for real applications in the space of Industrial IoT.