Federated Learning (FL) enables decentralized machine learning while preserving data privacy, making it ideal for sensitive applications where data cannot be shared. While FL has been widely studied in supervised contexts, its application to unsupervised learning remains underdeveloped. This work introduces FedCRef, a novel unsupervised federated learning method designed to uncover all underlying data distributions across decentralized clients without requiring labels. This task, known as Federated Clustering, presents challenges due to heterogeneous, non-uniform data distributions and the lack of centralized coordination. Unlike previous methods that assume a one-cluster-per-client setup or require prior knowledge of the number of clusters, FedCRef generalizes to multi-cluster-per-client scenarios. Clients iteratively refine their data partitions while discovering all distinct distributions in the system. The process combines local clustering, model exchange and evaluation via reconstruction error analysis, and collaborative refinement within federated groups of similar distributions to enhance clustering accuracy. Extensive evaluations on four public datasets (EMNIST, KMNIST, Fashion-MNIST and KMNIST49) show that FedCRef successfully identifies true global data distributions, achieving an average local accuracy of up to 95 %. The method is also robust to noisy conditions, scalable, and lightweight, making it suitable for resource-constrained edge devices.

Fully decentralised federated learning enables collaborative model training among edge devices without relying on a central coordinator, thereby avoiding single points of failure and supporting spontaneous collaboration in pervasive environments. However, the absence of coordination introduces challenges that go beyond data heterogeneity alone. In realistic decentralised settings, devices often start from different model initializations, possess limited and non-IID local data, and interact over unstructured communication graphs, making naive parameter averaging ineffective and potentially destructive. In this paper, we address decentralised learning under combined data and initial model heterogeneity by proposing DecDiff+VT, a coordination-free decentralised learning algorithm specifically designed for such environments. DecDiff+VT integrates two complementary mechanisms: DecDiff, a disruption-aware aggregation strategy that updates local models toward their neighborhood average with a magnitude inversely proportional to model disagreement, and a lightweight virtual teacher (VT) mechanism based on soft-label regularization to improve local generalization in the absence of strong or centralized teacher models. Extensive experiments on image classification and activity recognition benchmarks (MNIST, Fashion-MNIST, EMNIST, CIFAR-10, and UCI-HAR) show that DecDiff+VT consistently outperforms or matches state-of-the-art decentralised baselines, achieving faster convergence, improved generalization, and greater robustness to overfitting, without incurring additional communication or memory overhead compared to standard decentralised averaging.

Federated Learning (FL) has emerged as a privacypreserving paradigm that enables collaborative model training between distributed client devices without exchanging raw data. However, while FL mitigates direct data exposure, the influence of client data remains encoded in the trained model, raising concerns in scenarios where data deletion is legally or ethically required. Federated Unlearning (FU) aims to address this gap. Yet first existing FU solutions often rely on strong assumptions, such as maintaining complete histories of model updates or leveraging publicly available datasets that resemble private client data, or propose solutions which are not selective and coarsely reinitialize part of the learned model (with the associated non-negligible overhead). This work introduces a novel unlearning method based on distilled synthetic data, which clients can generate and transmit to the server; these compact and distilled samples are unrecognizable from the original data, but preserve most of their training influence, thus enabling targeted removal from the global model. The unlearning phase is executed on the server side, without requiring the target client to be online or maintaining any historical record of client participation. In the reported experimental results, we show that our method can achieve competitive or superior unlearning performance compared to state-of-the-art baselines, in particular in terms of more precise forgetting, using a very small distilled dataset (e.g., distilling only five data points per class).

Decentralized federated learning (DFL) enables devices to collaboratively train models over complex network topologies without relying on a central controller. In this setting, local data remains private, but its quality and quantity can vary significantly across nodes. The extent to which a fully decentralized system is vulnerable to poor-quality or corrupted data remains unclear, but several factors could contribute to potential risks. Without a central authority, there can be no unified mechanism to detect or correct errors, and each node operates with a localized view of the data distribution, making it difficult for the node to assess whether its perspective aligns with the true distribution. Moreover, models trained on low-quality data can propagate through the network, amplifying errors. To explore the impact of low-quality data on DFL, we simulate two scenarios with degraded data quality—one where the corrupted data is evenly distributed in a subset of nodes and one where it is concentrated on a single node—using a decentralized implementation of FedAvg. Our results reveal that averaging-based decentralized learning is remarkably robust to localized bad data, even when the corrupted data resides in the most influential nodes of the network. Counterintuitively, this robustness is further enhanced when the corrupted data is concentrated on a single node, regardless of its centrality in the communication network topology. This phenomenon is explained by the averaging process, which ensures that no single node—however central—can disproportionately influence the overall learning process.

In the active landscape of AI research, decentralised learning is gaining momentum. Decentralised learning allows individual nodes to keep data locally where they are generated and to share knowledge extracted from local data among themselves through an interactive process of collaborative refinement. This paradigm supports scenarios where data cannot leave the data owner node due to privacy or sovereignty reasons or real-time constraints imposing proximity of models to locations where inference has to be carried out. The distributed nature of decentralised learning implies significant new research challenges with respect to centralised learning. Among them, in this paper, we focus on robustness issues. Specifically, we study the effect of nodes’ disruption on the collective learning process. Assuming a given percentage of “central” nodes disappear from the network, we focus on different cases, characterised by (i) different distributions of data across nodes and (ii) different times when disruption occurs with respect to the start of the collaborative learning task. Through these configurations, we are able to show the non-trivial interplay between the properties of the network connecting nodes, the persistence of knowledge acquired collectively before disruption or lack thereof, and the effect of data availability pre- and post-disruption. Our results show that decentralised learning processes are remarkably robust to network disruption. As long as even minimum amounts of data remain available somewhere in the network, the learning process is able to recover from disruptions and achieve significant classification accuracy. This clearly varies depending on the remaining connectivity after disruption, but we show that even nodes that remain completely isolated can retain significant knowledge acquired before the disruption.

Fully decentralized learning is gaining momentum for training AI models at the Internet’s edge, addressing infrastructure challenges and privacy concerns. In a decentralized machine learning system, data is distributed across multiple nodes, with each node training a local model based on its respective dataset. The local models are then shared and combined to form a global model capable of making accurate predictions on new data. Our exploration focuses on how different types of network structures influence the spreading of knowledge - the process by which nodes incorporate insights gained from learning patterns in data available on other nodes across the network. Specifically, this study investigates the intricate interplay between network structure and learning performance using three network topologies and six data distribution methods. These methods consider different vertex properties, including degree centrality, betweenness centrality, and clustering coefficient, along with whether nodes exhibit high or low values of these metrics. Our findings underscore the significance of global centrality metrics (degree, betweenness) in correlating with learning performance, while local clustering proves less predictive. We highlight the challenges in transferring knowledge from peripheral to central nodes, attributed to a dilution effect during model aggregation. Additionally, we observe that central nodes exert a pull effect, facilitating the spread of knowledge. In examining degree distribution, hubs in Barabási–Albert networks positively impact learning for central nodes but exacerbate dilution when knowledge originates from peripheral nodes. Finally, we demonstrate the formidable challenge of knowledge circulation outside of segregated communities, and discuss the impact of class cross-correlations.

Federated learning (FL) is proving to be one of the most promising paradigms for leveraging distributed resources, enabling a set of clients to collaboratively train a machine learning model while keeping the data decentralized. The explosive growth of interest in the topic has led to rapid advancements in several core aspects like communication efficiency, handling non-IID data, privacy, and security capabilities. However, the majority of FL works only deal with supervised tasks, assuming that clients' training sets are labeled. To leverage the enormous unlabeled data on distributed edge devices, in this paper, we aim to extend the FL paradigm to unsupervised tasks by addressing the problem of anomaly detection (AD) in decentralized settings. In particular, we propose a novel method in which, through a preprocessing phase, clients are grouped into communities, each having similar majority (i.e., inlier) patterns. Subsequently, each community of clients trains the same anomaly detection model (i.e., autoencoders) in a federated fashion. The resulting model is then shared and used to detect anomalies within the clients of the same community that joined the corresponding federated process. Experiments show that our method is robust, and it can detect communities consistent with the ideal partitioning in which groups of clients having the same inlier patterns are known. Furthermore, the performance is significantly better than those in which clients train models exclusively on local data and comparable with federated models of ideal communities' partition.

The most widely adopted approach for knowledge extraction from raw data generated at the edges of the Internet (e.g., by IoT or personal mobile devices) is through global cloud platforms, where data is collected from devices, and analysed. However, with the increasing number of devices spread in the physical environment, this approach rises several concerns. The data gravity concept, one of the basis of Fog and Mobile Edge Computing, points towards a decentralisation of computation for data analysis, whereby the latter is performed closer to where data is generated, for both scalability and privacy reasons. Hence, data produced by devices might be processed according to one of the following approaches: (i) directly on devices that collected it (ii) in the cloud, or (iii) through fog/mobile edge computing techniques, i.e., at intermediate nodes in the network, running distributed analytics after collecting subsets of the data. Clearly, (i) and (ii) are the two extreme cases of (iii). It is worth noting that the same analytics task executed at different collection points in the network, comes at different costs in terms of traffic generated over the network. Precisely, these costs refer to the traffic generated to move data towards the collection point selected (e.g. the Edge or the Cloud) and the one induced by the distributed analytics process. Until now, deciding if to use intermediate collection points, and which one they should be in order to both obtain a target accuracy and minimise the network traffic, is an open question. In this paper, we propose an analytical framework able to cope with this problem. Precisely, we consider learning tasks, and define a model linking the accuracy of the learning task performed with a certain set of collection points, with the corresponding network traffic. The model can be used to identify, given the specification of the learning problem (e.g. binary classification, regression, etc.), and its target accuracy, what is the optimal level for collecting data in order to minimise the total network cost. We validate our model through simulations in order to show that setting, in simulation, the level of intermediate collection indicated by our model, leads to the minimum cost for the target accuracy.

It is commonly assumed that in a smart city there will be thousands of mostly mobile/wireless smart devices (e.g. sensors, smart-phones, etc.) that will continuously generate big amounts of data. Data will have to be collected and processed in order to extract knowledge out of it, to feed users' and smart city applications. A typical approach to process such big amounts of data is to i) gather all the collected data on the cloud through wireless pervasive networks, and ii) perform data analysis operations exploiting machine learning techniques. However, according to many studies, this centralised cloud-based approach may not be sustainable from a networking point of view. The joint effect of data-intensive users' multimedia applications and smart cities monitoring and control applications may result in severe network congestions making applications hardly usable. To cope with this problem, in this paper we propose a distributed machine learning approach that does not require to move data in a centralised cloud platform, but processes it directly where it is collected. Specifically, we exploit Hypothesis Transfer Learning (HTL) to build a distributed machine learning framework. In our framework we train a series of partial models, each ''residing'' in a location where a subset of the dataset is generated. We then refine the partial models by exchanging them between locations, thus obtaining a unique complete model. Using an activity classification task on a reference dataset as a concrete example, we show that the classification accuracy of the HTL model is comparable with that of a model built out of the complete dataset, but the cost in term of network overhead is dramatically reduced. We then perform a sensitiveness analysis to characterise how the overhead depends on key parameters. It is also worth noticing that the HTL approach is suitable for applications dealing with privacy sensitive data, as data can stay where they are generated, and do not need to be transferred to third parties, i.e., to a cloud provider, to extract knowledge out of it.

In cyber-physical convergence scenarios information flows seamlessly between the physical and the cyber worlds. Here, users’ mobile devices represent a natural bridge through which users process acquired information and perform actions. The sheer amount of data available in this context calls for novel, autonomous and lightweight data-filtering solutions, where only relevant information is finally presented to users. Moreover, in many real-world scenarios data is not categorised in predefined topics, but it is generally accompanied by semantic descriptions possibly describing users’ interests. In these complex conditions, user devices should autonomously become aware not only of the existence of data in the network, but also of their semantic descriptions and correlations between them. To tackle these issues, we present a set of algorithms for knowledge and data dissemination in opportunistic networks, based on simple and very effective models (called cognitive heuristics) coming from cognitive sciences. We show how to exploit them to disseminate both semantic data and the corresponding data items. We provide a thorough performance analysis, under various different conditions comparing our results against non-cognitive solutions. Simulation results demonstrate the superior performance of our solution towards a more effective semantic knowledge acquisition and representation, and a more tailored content acquisition.

Opportunistic networking is one of the key paradigms to support direct communication between devices in a mobile scenario. In this context, the high volatility and dynamicity of information and the fact that mobile nodes have to make decisions in condition of partial or incomplete knowledge, makes the development of effective and efficient data dissemination schemes very challenging. In this paper we present algorithms based on well-established models in cognitive sciences, in order to disseminate both data items, and semantic information associated with them. In our approach, semantic information represents both meta-data associated to data items (e.g., tags associated to them), and meta-data describing the interests of the users (e.g., topics for which they would like to receive data items). Our solution exploits dissemination of semantic data about the users' interests to guide the dissemination of the corresponding data items. Both dissemination processes are based on models coming from the cognitive sciences field, named cognitive heuristics, which describe how humans organise information in their memory and exchange it during interactions based on partial and incomplete information. We exploit a model describing how semantic data can be organised in each node in a semantic network, based on how humans organise information in their memory. Then, we define algorithms based on cognitive heuristics to disseminate both semantic data and data items between nodes upon encounters. Finally, we provide initial performance results about the diffusion of interests among users, and the corresponding diffusion of data items.

Opportunistic Networks (OppNets) offer a very volatile and dynamic networking environment. Several applications proposed for OppNets - such as social networking, emergency management, pervasive and urban sensing - involve the problem of sharing content amongst interested users. Despite the fact that nodes have limited resources, existing solutions for content sharing require that the nodes maintain and exchange large amount of status information, but this limits the system scalability. In order to cope with this problem, in this paper we present and evaluate a solution based on cognitive heuristics. Cognitive heuristics are functional models of the mental processes, studied in the cognitive psychology field. They describe the behavior of the brain when decisions have to be taken quickly, in spite of incomplete information. In our solution, nodes maintain an aggregated information built up from observations of the encountered nodes. The aggregate status and a probabilistic decision process is the basis on which nodes apply cognitive heuristics to decide how to disseminate content items upon meeting with each other. These two features allow the proposed solution to drastically limit the state kept by each node, and to dynamically adapt to both the dynamics of item diffusion and the dynamically changing node interests. The performance of our solution is evaluated through simulation and compared with other solutions in the literature.

We introduce a new paradigm of neural networks where neurons autonomously search for the best reciprocal position in a topological space so as to exchange information more profitably. The idea that elementary processors move within a network to get a proper position is borne out by biological neurons in brain morphogenesis. The basic rule we state for this dynamics is that a neuron is attracted by the mates which are most informative and repelled by ones which are most similar to it. By embedding this rule into a Newtonian dynamics, we obtain a network which autonomously organizes its layout. Thanks to this further adaptation, the network proves to be robustly trainable through an extended version of the back-propagation algorithm even in the case of deep architectures. We test this network on two classic benchmarks and thereby get many insights on how the network behaves, and when and why it succeeds.

We introduce a wait-and-chase scheme that models the contact times between moving agents within a connectionist construct. The idea that elementary processors move within a network to get a proper position is borne out both by biological neurons in the brain morphogenesis and by agents within social networks. From the former, we take inspiration to devise a medium-term project for new artificial neural network training procedures where mobile neurons exchange data only when they are close to one another in a proper space (are in contact). From the latter, we accumulate mobility tracks experience. We focus on the preliminary step of characterizing the elapsed time between neuron contacts, which results from a spatial process fitting in the family of random processes with memory, where chasing neurons are stochastically driven by the goal of hitting target neurons. Thus, we add an unprecedented mobility model to the literature in the field, introducing a distribution law of the intercontact times that merges features of both negative exponential and Pareto distribution laws. We give a constructive description and implementation of our model, as well as a short analytical form whose parameters are suitably estimated in terms of confidence intervals from experimental data. Numerical experiments show the model and related inference tools to be sufficiently robust to cope with two main requisites for its exploitation in a neural network: the nonindependence of the observed intercontact times and the feasibility of the model inversion problem to infer suitable mobility parameters.