Blog

This work was conducted by researchers João Victor Feio Gonçalves (UFPA), John Sousa (UFPA), Rafael Veiga (UFPA), Lucas de Lima Bastos (UNIFESSPA), Lucas de Sousa Pacheco (UFPA), Iago Medeiros (UFPA), Denis Rosário (UFPA), and Eduardo Coelho Cerqueira (UFPA).

They developed ASTRA, a new method designed to optimize Federated Learning in scenarios where data across devices is highly diverse and disconnected. The core of the innovation lies in Dual-Space Regularization, which combines physical constraints on model weights with semantic self-distillation to ensure that local predictions remain aligned with the global objective.

To enable use on resource-constrained devices, the authors implement Curriculum-Based Scaling, activating the “instructor” model intensively only at the beginning and subsequently on a periodic basis. The purpose of the study is to mitigate client drift and prevent catastrophic system divergence, achieving higher accuracy than traditional methods with only a small increase in computational cost.

The project developed by Amanda Lopes (UFPA), John Sousa (UFPA), Lucas de Lima Bastos (UNIFESSPA), Lucas de Sousa Pacheco (UFPA), Iago Medeiros (UFPA), Denis Rosário (UFPA), and Eduardo Coelho Cerqueira (UFPA) was presented at SBRC 2026.

Vehicles are increasingly capable of collecting large amounts of data through sensors and IoT devices. In the context of autonomous driving, one approach is to equip cars with multiple sensors and sufficient computing power to process this information and estimate, in real time, the best route to take.

With this processing, the vehicle itself can make decisions, such as braking, accelerating, or changing its course. However, the potential is even greater when different vehicles can participate in the same network and share information. Based on the data and predictions generated by these vehicles, it is possible, for example, to identify the best routes and anticipate traffic situations.

It is in this context that connected autonomous vehicles have emerged: an attempt to expand the capabilities of vehicle systems to solve problems that the hardware available in a single vehicle cannot solve on its own. The idea is to leverage the knowledge generated by academia and create solutions that can, in the future, be implemented by the industry.

One of the challenges is enabling the vehicle to “read” its surrounding context and use that information to make more appropriate decisions. This is where PRINCE comes in—a federated learning (FL) architecture designed for autonomous vehicle networks.

The proposal aims to select, from among the available vehicles, those best suited to participate in a federated learning process. Instead of considering all vehicles in the network indiscriminately, the system evaluates different characteristics to determine which ones can contribute most significantly to the learning process.

Proactive customer selection, as exemplified by PRINCE, offers significant advantages over traditional reactive methods in Connected and Autonomous Vehicles (CAVs) environments, particularly in terms of resource efficiency, convergence speed, and model stability. While reactive approaches select vehicles and address any failures only after they occur, the proactive strategy seeks to anticipate potential problems and select, from the outset, the clients most likely to complete training.

One of the main benefits lies in the efficient use of resources. In reactive methods, bandwidth and processing capacity can be wasted when selected vehicles drop out or fail during a training round. The proactive approach seeks to pre-filter unstable nodes, achieving—according to the presented results—an improvement of up to 3.5 times in resource efficiency. This proactive approach also contributes to a significant reduction in connection failures—up to 2.2 to 2.5 times fewer compared to reactive approaches—thanks to the use of stochastic models to predict instabilities and pre-filtering based on signals such as RSSI.

These gains are also reflected in the model’s performance. PRINCE achieved a final accuracy of 83.90%, outperforming state-of-the-art reactive models by up to 9.22%. Furthermore, proactive selection accelerates convergence: in one of the evaluated scenarios, the system reached 60% accuracy three times faster than random selection methods, while also outperforming reactive approaches such as ACSP-FL and ECS-HDSR.

Another key advantage lies in reduced latency and computational costs. In reactive strategies, such as ECS-HDSR, when a client fails, the system must find a replacement and restart or reorganize part of the process, which adds steps, time, and processing load. The proactive approach seeks to avoid this situation by pre-selecting a set of clients with a high probability of remaining connected and completing the training.

This proactive approach also promotes training stability. While reactive methods may exhibit fluctuations in accuracy and spikes in the loss function due to the inclusion of models from failed clients, proactive selection tends to produce a more stable and consistent learning trajectory.

In summary, the fundamental difference lies in how each strategy handles uncertainty. Reactive selection waits for a problem to occur before attempting to correct it; proactive selection uses vehicle behavior history, network conditions, and mobility forecasts to anticipate potential failures and prevent them. In a dynamic environment such as that of connected and autonomous vehicles, this predictive capability can make each communication cycle more efficient, reduce waste, and contribute to global models that converge more quickly and stably.

Factors Considered in the Active Selection of Vehicles

The first factor considered is the location and connectivity of the vehicles, which are determined by the Received Signal Strength Indicator (RSSI). Vehicles that are closer or farther away—and whether or not they have internet coverage—are evaluated. Vehicles with weak RSSI values (i.e., values below -85 dbM) are excluded from further selection, as they do not have sufficient connectivity in that area to actively participate in the federated learning client selection process. Vehicles with good RSSI values (i.e., values above -85 dbM) are selected for the next stage.

To estimate the likelihood that a vehicle will remain connected during the process, in the second step the architecture uses the concept of a Markov chain—a stochastic process in which the system can be in a given state and either remain in that state or transition to another. Based on this model, the probability that a given vehicle will remain connected to the network is calculated.

This information is important because, in a federated learning process, it is not enough for a vehicle to have relevant data: it must also be able to remain connected for the time necessary to participate in the training.

The Importance of Data Heterogeneity

The third and final criterion used by PRINCE is the diversity of the data available in each vehicle. To this end, the architecture takes into account Shannon entropy, a measure related to the variability and diversity of information.

The reasoning is that vehicles with more diverse data can make a more meaningful contribution to the learning process, since they allow the model to encounter different situations and patterns.

Thus, the system seeks to prioritize vehicles with a higher entropy index—that is, greater diversity in their data. The architecture therefore combines different criteria to select the most suitable vehicles to participate in federated learning. Generally speaking, three dimensions are considered: proximity—prioritizing vehicles that are close to the region or situation of interest; connectivity—selecting vehicles most likely to remain connected during the process; and data heterogeneity—prioritizing vehicles whose data exhibit greater diversity.

The goal is to find a set of vehicles that is not only related to the problem we want to analyze but is also capable of effectively contributing to the model’s training.

Possible Applications

One possible application would be the use of PRINCE by a company that needs to transmit specific information to certain vehicles in a given region.

Imagine, for example, that a problem is occurring on a particular route. The company could use the architecture to identify vehicles that are in that area or heading toward it and share relevant information with those users.

Similarly, a driver traveling to a specific city could receive specific information related to conditions in that region, allowing the system to use the data generated by connected vehicles to support more informed decisions.

Read the full article in the SBRC 2026 proceedings 

The paper, authored by Rodrigo Almeida Bezerra (UFBA), Luiz Fernando Bittencourt (UNICAMP), Leandro Villas (UNICAMP), and Maycon Peixoto (UFBA), is set within the intersection of the fields of IoT and Brain-Computer Interfaces (BCI).

The study simulates the operation of a BCI. In this specific case, the BCI involves a game that connects via electroencephalography—a device that, when placed on the head, captures brain signals to establish a connection between the user’s brain and the game.

In this context, potential delays can negatively impact the user experience. Edge computing comes into play, then, to enable a fast, low-latency user experience. To achieve this, processing must take place at the edge; that is why the shift from the cloud to the edge is important for specific applications like this one.

In the case of this work, there is an additional dimension: user mobility. The user is moving from one place to another, and the application modules need to move along with the user. User mobility, therefore, complicates the interactions between the system’s elements; the edge could be, for example, a device in the user’s home, and when the user moves away, latency increases.

It is in this context that the MX algorithm solution comes into play, to track the user.

What is the MI-X algorithm, and what is its main objective?

This algorithm, developed by the researchers, is a strategy for migrating game applications from the fog to the cloud.

Mobility Impact-X (MI-X) is a dynamic allocation algorithm and a mobility-aware modular allocation strategy designed to operate in Hierarchical Fog Computing architectures. It was specifically designed to manage modular, latency-sensitive Brain-Computer Interface (BCI) applications, such as electroencephalogram-based games (EEG Games), within the cloud-fog-edge continuum.

Its primary goal is to minimize application latency and response times while simultaneously optimizing the use of computational resources. To achieve this, MI-X seeks to dynamically balance the following factors:

Computational proximity to the user: Prioritizing execution at the edge (cloudlets) to reduce perceived latency.

Communication cost: Assessing the impact of distributing or migrating interdependent modules across the network.

Operational stability: Avoiding abrupt migrations and long hierarchical hops that could cause system instability.

What sets MI-X apart is its ability to integrate the application’s modular structure with mobility management. This allows the algorithm to keep interdependent components physically close to one another within the infrastructure, preventing network costs from negating individual processing gains and ensuring consistent performance even amid continuous user mobility.

Challenges

Edge applications are close to the user but have limited resources; cloud applications, on the other hand, have more resources but are farther from the user, resulting in longer response times.

In urban mobility scenarios, edge resources can become saturated. The developed solution analyzes the impact of each option: it performs a search and attempts to balance latency, computational capacity, communication costs, etc.

An interesting fact about the development of this solution is that the history of this algorithm includes others, built in 2019, 2020, and 2023, which have evolved over time; MI-X is the first in the series to offer the possibility of reducing latency.

The researchers emphasize that it is not enough to simply place everything at the edge; processing must intelligently follow the user, tracking their movements.

Work continues to evolve through the research of master’s student Rodrigo Bezerra (UFBA)—now with a new strategy called HERMES (Heuristic Resource Management for Edge Systems). This approach builds on MI-X by considering not only where application modules should be executed as the user moves, but also when a migration is actually worthwhile. HERMES analyzes the impact of each possible reallocation, taking into account latency, resource availability, and the cost of transferring modules, thereby avoiding unnecessary migrations and the so-called “back-and-forth” effect between the edge, fog, and cloud. Thus, the research evolves from the idea of simply following the user to that of intelligently and stably orchestrating the entire path between the edge, the fog, and the cloud.

Read the paper 

The paper, authored by ICoNIoT researchers João V. D. Sobrinho (UFRJ) and Miguel E. M. Campista (UFRJ), was presented at the 18th Brazilian Symposium on Ubiquitous and Pervasive Computing at CSBC 2026.
The research grew out of a project related to the automotive industry and investigates an important problem for Artificial Intelligence applications: how to train computer vision models when there is little labeled data available?

In supervised learning, the machine needs examples accompanied by labels—something similar to an answer key. The model receives an image, produces a response, compares that response with the correct label, and, based on the error, adjusts its behavior. The better and more numerous the examples used in training, the greater the model’s ability to perform the task tends to be.

In the project that gave rise to this work, the goal is to use federated learning in the automotive context, particularly due to data privacy concerns. One of the applications being considered is the identification of VRUs (Vulnerable Road Users)—such as pedestrians, cyclists, and skateboarders—so that a system can recognize them and issue alerts to drivers.

For this system to work, however, the model needs to learn how to identify these users in the images. In this context, there is a fundamental question: who will tell the machine what is in each image?

Imagine that thousands of images of streets and roads are collected. To use them for training, someone would need to identify and label, one by one, the presence of pedestrians, cyclists, and other elements. This labeling process is labor-intensive, expensive, and not very scalable. Furthermore, in a real-world scenario, it would be unreasonable to expect the client or end user to perform this work themselves.

One alternative is to use a pre-trained model to automatically generate these labels. The problem is that this model can also make many errors. If the generated labels are incorrect, these errors end up being incorporated into the training and can compromise the performance of the final model.

This is where SAM3 comes in, which is used in this study to generate pseudo-labels. Instead of relying solely on a person to manually identify objects in the images, the model is used to automatically generate these labels. The goal is to evaluate whether these pseudo-labels can improve the quality of the available data and, consequently, the performance of the object detection system in a scenario where labels are scarce.
The term “pseudo-label” stems precisely from the fact that the label was not produced directly by a human, but by the Artificial Intelligence model itself.

There is, however, an important privacy-related issue. SAM3 is a Meta model, and when images are sent to an external solution, concerns arise about what happens to that data. In this work, SAM3 runs locally, reducing privacy risks by avoiding the need to send these images to an external service.

From Research to the Real World

One of the points raised by Miguel Elias M. Campista is that the issue of labeling in federated learning has yet to be fully explored from a practical perspective. There is a significant amount of research on models and methods, but another question remains unanswered: how can these solutions be transformed into systems that can effectively function in the real world?

When a technology moves beyond the controlled environment of academic research and is incorporated into a product or service, various challenges arise that do not always appear in scientific experiments—or that become even more pronounced—such as privacy, data quality, implementation costs, infrastructure, scalability, and interaction with real users.
Thus, this work is not merely about finding a way to improve object detection. It also addresses a broader issue: how to bridge the gap between knowledge produced in academia and its implementation in real-world applications.

Read the full article in the CSBC proceedings 

The presentation is entilted “Intelligence Across the IoT Continuum: Integration, Adaptation, and Distributed Processing”

The increasing integration of artificial intelligence into Internet of Things (IoT) systems has intensified the challenges of integration, data processing, and decision-making in distributed environments. In this context, the computing continuum enables the distribution of data and services across devices, edge, fog, and cloud infrastructures according to the available resources and the requirements of each application. Drawing on research conducted in the domain of smart agriculture, this presentation discusses architectures for heterogeneous platform integration, dynamic adaptation, service placement, and intelligent data processing. The results highlight the trade-offs involved in selecting where services should be executed and emphasize the importance of considering connectivity, latency, computational capacity, and resilience. This perspective contributes to the development of more autonomous, adaptive, and AI-ready IoT systems.

The speaker

Dener Edson Ottolini Guedes da Silva holds a Ph.D. in Information Engineering and an M.Sc. in Computer Science from the Federal University of ABC (UFABC), as well as a degree in Information Technology for Business Management. He is currently a postdoctoral researcher at FEI University Center, where he investigates the development and deployment of intelligent applications across the computing continuum. His research interests include the Internet of Things, distributed systems, edge, fog, and cloud computing, programmable networks with P4, artificial intelligence, and smart agriculture.

The project, led by researchers João Silva e Costa (GTA-PEE/COPPE-DEL/Poli, Federal University of Rio de Janeiro – UFRJ), Guilherme A. Thomaz, Ronaldo A. Ferreira, and Miguel Elias M. Campista (Facom, Federal University of Mato Grosso do Sul – UFMS), was presented at the SBRC 2026

FLeer2FLeer (F2F) is a web-based tool for orchestrating federated learning based on a P2P architecture. F2F introduces an indexer responsible for discovering, announcing, and enrolling clients in multiple federations, without interfering with the training process.

The tool provides a decentralized orchestration infrastructure for Federated Learning (FL), enabling the training of artificial intelligence models without compromising the privacy of the users who own the data.

The project, led by researchers João Silva e Costa (GTA-PEE/COPPE-DEL/Poli, Federal University of Rio de Janeiro – UFRJ), Guilherme A. Thomaz, Ronaldo A. Ferreira, and Miguel Elias M. Campista (Facom, Federal University of Mato Grosso do Sul – UFMS), was presented at the SBRC 2026 Tools Exhibition and is available in the symposium proceedings.

Federated Learning

This work relates to federated learning, an approach initially developed by Google to address a central challenge in training artificial intelligence models: the need to collect large volumes of user data, which raised privacy concerns.

Instead of centralizing user data on a single server to train models, the federated learning approach reverses this logic. The model is sent to the devices or organizations that hold the data, where training is performed locally. Subsequently, only updates to the trained models are sent back to the central server, which combines (“merges”) these contributions to produce an increasingly accurate and convergent global model. This process depends on the participation of multiple clients training the model in parallel.

The procedure is repeated until the model converges or a stopping criterion is met. Because this distributed training approach preserves user privacy, it is well-suited for situations where training requires sensitive user information, such as data from Internet of Things (IoT) systems, medical data, or data generated by mobile devices.

Peer-to-Peer

Much of the existing literature assumes that these training processes are already underway. However, there is little discussion of a practical question: How does a new participant join a federated training process? How can a company or a user who possesses data find a compatible training process and begin to contribute?

It is precisely this gap that the article seeks to address. The proposal allows a client—whether a company or an end user who owns the data—to identify federated learning models that align with its interests and participate in the corresponding training processes.

By contributing their own data, customers help improve the model and, at the same time, obtain a system that is better tailored to their specific needs. In addition, existing models can continue to be refined on an ongoing basis as new data is incorporated into the training process.

Proposal

This research proposes the development of a web-based tool inspired by P2P architecture, called FLeer2FLeer (F2F), for training federated models. F2F proposes a server that acts as an indexer for discovering and managing federations. Unlike approaches focused on complex changes to the algorithm, the tool prioritizes the user experience and the governance of multiple simultaneous model training sessions.

FLeer2FLeer is an innovative tool that combines federated learning and peer-to-peer architecture to optimize the training of artificial intelligence models. Unlike traditional methods that centralize private information, this solution allows data to remain with the users, promoting greater privacy and security during processing.

The system facilitates collaboration between companies and individuals, enabling interested parties to locate and contribute to specific training programs in a distributed manner. By decentralizing the computational effort, the platform helps reduce operational costs and makes the models more accurate and tailored to the participants’ actual needs. The central focus is on overcoming the challenges of practical implementation by efficiently connecting data owners to processes for orchestrating global models.

Although the system is not an open public service, it can be deployed within companies, allowing different organizations to conduct collaborative training on distributed data without the need to centralize it.

This approach also makes it possible to distribute the computational cost among participants, since processing is carried out in a decentralized manner. The main challenge now is to transform this solution into a technology that can be applied in real-world scenarios, expanding its adoption beyond the research environment.

The conference is one of the most important in the field, and this year Brazilians will have the chance to participate without leaving the country

The IEEE/ACM International Conference on Big Data Computing, Applications, and Technologies (BDCAT 2026) will take place in December 2026 in Florianópolis. Two researchers from ICoNIoT are serving on the conference committees: Luiz Bittencourt (Unicamp) is one of the General Chairs, and Flávia Delicato (UFF) is one of the Program Chairs. Learn more

The Conference

As the era of data-driven intelligence continues to reshape our world, BDCAT remains a premier annual conference series dedicated to pushing the boundaries of Big Data. Its mission is to provide a dynamic platform for researchers, professionals, and industry leaders from around the world to present new findings, share innovative ideas, and tackle the most pressing challenges in the broad field of Big Data computing and applications.

Since its inaugural edition in London, United Kingdom (2014), BDCAT has grown into a vibrant global community. Over the past decade, the conference has traveled to the world’s leading scientific hubs and has played a key role in bridging the gap between academia and industry. By promoting interdisciplinary collaboration, the BDCAT community has directly impacted how we scale, manage, and extract useful insights from massive datasets, influencing everything from business systems and scientific research to public policy and deep learning architectures.

Deadlines and Dates

The conference will take place from December 1 to 4, 2026, and the call for papers is open until August 19.

Full details are available on the conference website

UCC

Running concurrently with BDCAT is the UCC Conference—the International Conference on Utility and Cloud Computing—also taking place from December 1 to 4. The technical committee is coordinated by ICoNIoT researcher Edmundo Madeira.

As the digital landscape increasingly relies on decentralized, scalable, and on-demand computing resources, UCC stands out as a premier international forum for discussing the present and future of utility, cloud, and cutting-edge computing paradigms. UCC 2026 brings together prominent researchers from research institutions and industry to present pioneering research, architectural innovations, and practical applications that are driving the next generation of cloud technologies.

The conference is one of the most important in the field, and this year Brazilians will have the chance to participate without leaving the country

The IEEE/ACM International Conference on Big Data Computing, Applications, and Technologies (BDCAT 2026) will take place in December 2026 in Florianópolis. Two researchers from ICoNIoT are serving on the conference committees: Luiz Bittencourt (Unicamp) is one of the General Chairs, and Flavia Delicato (UFF) is one of the Program Chairs. Learn more

The Conference

As the era of data-driven intelligence continues to reshape our world, BDCAT remains a premier annual conference series dedicated to pushing the boundaries of Big Data. Its mission is to provide a dynamic platform for researchers, professionals, and industry leaders from around the world to present new findings, share innovative ideas, and tackle the most pressing challenges in the broad field of computing and Big Data applications.

Since its inaugural edition in London, United Kingdom (2014), BDCAT has grown into a vibrant global community. Over the past decade, the conference has traveled to the world’s leading scientific hubs and has played a key role in bridging the gap between academia and industry. By promoting interdisciplinary collaboration, the BDCAT community has directly impacted the way we scale, manage, and extract useful insights from massive datasets, influencing everything from enterprise systems and scientific research to public policy and deep learning architectures.

Deadlines and Dates

The conference will take place from December 1 to 4, 2026, and the call for papers is open until August 19.

Full details are available on the conference website

The conference is one of the most important in the field, and this year Brazilians will have the chance to participate without leaving the country

The IEEE/ACM International Conference on Big Data Computing, Applications, and Technologies (BDCAT 2026) will take place in December 2026 in Florianópolis. Two researchers from ICoNIoT are serving on the conference committees: Luiz Bittencourt (Unicamp) is one of the General Chairs, and Flávia Delicato (UFF) is one of the Program Chairs. Learn more

The Conference

As the era of data-driven intelligence continues to reshape our world, BDCAT remains a premier annual conference series dedicated to pushing the boundaries of Big Data. Its mission is to provide a dynamic platform for researchers, professionals, and industry leaders from around the world to present new findings, share innovative ideas, and tackle the most pressing challenges in the broad field of Big Data computing and applications.

Since its inaugural edition in London, United Kingdom (2014), BDCAT has grown into a vibrant global community. Over the past decade, the conference has traveled to the world’s leading scientific hubs and has played a key role in bridging the gap between academia and industry. By promoting interdisciplinary collaboration, the BDCAT community has directly impacted how we scale, manage, and extract useful insights from massive datasets, influencing everything from business systems and scientific research to public policy and deep learning architectures.

Deadlines and Dates

The conference will take place from December 1 to 4, 2026, and the call for papers is open until August 19.

Full details are available on the conference website