Discovering Collaboration from Novelty: Random Network Distillation for Clustered Federated Learning Davide Domini
Gianluca Aguzzi
Ivana Dusparic
Danilo Pianini
Mirko Viroli
arXiv:2606.30499v1 [cs.LG] 29 Jun 2026
University of Bologna Trinity College Dublin University of Bologna University of Bologna University of Bologna Cesena, Italy Cesena, Italy Dublin, Ireland Cesena, Italy Cesena, Italy [email protected] [email protected] [email protected] [email protected] [email protected] 0000-0002-1553-4561 0000-0003-0621-5400 0000-0002-8392-5409 0000-0003-2702-5702 0009-0006-8337-8990
Abstract—Federated Learning often suffers under nonindependently and identically distributed data, where a single global model may fail to represent the diversity of client distributions. Clustered Federated Learning mitigates this issue by training specialized models for groups of similar clients, but existing approaches often couple cluster assignment with the main training loop, increasing computational and communication costs. We propose a lightweight clustering approach based on Random Network Distillation. Each client trains a compact Random Network Distillation predictor on its local data and uses its prediction error as a novelty signal to estimate similarity with other clients. This enables the discovery of meaningful client groups before federated training, without sharing raw data or repeatedly evaluating the main model. Crucially, the resulting federations emerge from local novelty estimates at runtime, making the method suitable for autonomous large-scale distributed systems where neither the number of clusters nor the collaboration structure can be specified a priori. Overall, by decoupling clustering from learning, the method provides a taskagnostic and efficient mechanism for autonomous collaboration under non-independently and identically distributed data. Index Terms—‘Federated Learning, Random Network Distillation, Clustering
I. I NTRODUCTION Federated Learning (FL) [1] has emerged as a prominent paradigm for collaborative model training in privacy-sensitive distributed settings, enabling devices to jointly learn shared models without exposing raw local data. While FL achieves strong performance under homogeneous data conditions, realworld deployments frequently exhibit non-independently and identically distributed (non-IID) data across devices, leading to degradation in model accuracy and convergence [2], [3]. This challenge becomes particularly severe in large-scale distributed systems comprising numerous devices that must learn and adapt over time. In many settings, data heterogeneity is not random but structured. Spatially distributed systems provide a clear example: devices in spatial proximity tend to observe similar phenomena, while devices in different regions exhibit statistically divergent distributions [4]. This naturally gives rise to group-level non-IID patterns, where clusters of devices share similar local distributions internally, while
distributions differ significantly across clusters. Furthermore, in such autonomous settings, manually specifying who should collaborate with whom or how many collaboration groups should exist is often unrealistic, since the number and structure of latent clusters are typically difficult to know in advance. Clustered Federated Learning (CFL) [5] addresses this by partitioning devices into groups and training a specialized model per cluster, rather than a single global model. Algorithms such as IFCA [6] and PSFL [7] have demonstrated that cluster-aware training substantially improves accuracy under non-IID conditions. However, these methods share a fundamental limitation: at each round, cluster assignment requires evaluating one or more instances of the full model being trained. In IFCA, every device must evaluate all cluster models on its local data and select the one yielding the lowest loss; in PSFL, devices continuously exchange and evaluate neighbors’ models to maintain coherent federations. This per-round evaluation is computationally expensive, communication-intensive, and may be prohibitive on resource-constrained devices. In this paper, we argue that clustering need not be entangled with the main learning loop. In many practical scenarios, the cluster structure is either stable or changes slowly relative to the learning timescale. It is therefore sufficient to perform clustering once (or periodically) using lightweight auxiliary models, entirely decoupled from the primary learning task. Toward this goal, we propose leveraging Random Network Distillation (RND) [8] as a mechanism for estimating data uncertainty and computing inter-device similarity for clustering. RND was originally introduced to quantify epistemic uncertainty in Multi-Agent Reinforcement Learning (MARL) for exploration and Transfer Learning (TL) [9], where the prediction error of a randomly initialized fixed network serves as a proxy for novelty with respect to an agent’s experience. We find this formulation remarkably well-suited to the federated clustering problem: devices can train small RND models on their local data and exchange them with neighbors to estimate distributional divergence–without sharing any raw data—enabling a lightweight, privacy-preserving clustering phase.
Concretely, our approach introduces a modular, plug-andplay pre-clustering phase applicable to existing FL algorithms. Devices first train compact RND networks (significantly smaller than the primary task model), exchange them within the network, and use the resulting uncertainty estimates to identify cluster membership. The primary learning phase then proceeds within the discovered clusters using any standard aggregation algorithm such as FedAvg [10]. When data distributions shift over time, the clustering phase can be retriggered periodically. To the best of our knowledge, this is the first work to apply RND for cluster discovery in FL. We validate our approach on well-known computer vision benchmarks augmented with synthetic noise to simulate feature skewness, a realistic and underexplored form of non-IID heterogeneity, demonstrating that RND uncertainty estimates effectively separate devices into meaningful clusters prior to any model training. II. BACKGROUND AND R ELATED W ORKS A. Federated Learning Federated Learning (FL) [1] is a distributed learning paradigm in which multiple clients collaboratively train machine learning models while keeping their raw data locally stored. In the standard server-based formulation, clients perform local optimization on private datasets and periodically send model updates to a central server, which aggregates them to produce an updated global model. The most common aggregation strategy is FedAvg [10], where the global model is obtained by averaging the parameters or updates received from the participating clients. While standard FL is effective when client data are sufficiently homogeneous, its performance can degrade under nonindependently and identically distributed (non-IID) data. In such settings, local updates may be biased toward different client-specific objectives, making their aggregation less stable and reducing the quality of the resulting global model [3], [5]. This issue has motivated several extensions of FL aimed at improving robustness under statistical heterogeneity. Clustered Federated Learning (CFL) addresses data heterogeneity by replacing the single global model with multiple specialized models. Clients are partitioned into clusters according to some notion of similarity, and each cluster trains a separate federated model. This formulation is particularly useful when the client population is composed of groups with internally similar but mutually different data distributions. Representative CFL approaches include IFCA [6], which alternates between assigning clients to the model that best fits their local data and updating cluster-specific models, and proximity-aware approaches such as PSFL [7], where client federations are formed and maintained according to local similarity estimates among neighboring devices. B. Data Heterogeneity Data heterogeneity is one of the main challenges in FL [11], since clients usually collect data through different devices, sensors, users, or environmental conditions rather than sampling
IID
II
No
D
n-
Fig. 1. Clustered data heterogeneity. Each node represents a device, and each background region denotes a different underlying feature distribution. Devices within the same cluster collect data from the same distribution, whereas devices across different clusters observe non-IID data.
from a single shared distribution. In real-world deployments, two clients may observe the same task from different perspectives: for instance, images of the same classes may be captured under different lighting conditions, cameras, backgrounds, styles, or noise patterns. Similarly, sensing devices deployed in different environments may measure the same type of phenomenon while being exposed to different local dynamics. Such heterogeneity can manifest in multiple forms, depending on how data are partitioned among clients [2], [3]. Commonly studied types of data skew include: (i) feature skew, where all clients share the same label space but differ in their feature distributions (e.g., handwritten character recognition with varying writing styles); (ii) label skew, in which each client observes only a subset of the global classes; and (iii) quantity skew, where clients possess highly imbalanced amounts of local data. Feature skew is therefore especially relevant when the semantic task is shared across clients, but the way data are observed varies across devices or contexts. As discussed in Section I, data skewness in FL can be structured rather than arbitrary. In this work, we formalize this setting as clustered feature skew, where clients belong to latent clusters with aligned feature distributions within each cluster and distinct distributions across clusters, as illustrated in Figure 1. Definition 1 (Clustered Feature Skew). Let D = {d1 , . . . , dN } be a set of clients partitioned into k latent clusters C1 , . . . , Ck . Each client di is associated with a local data distribution Pi (x, y), where x is the input and y is the corresponding label. We say that the clients satisfy clustered feature skew if all clients share the same label marginal distribution P (y), while their class-conditional feature distributions are identical within each cluster and different across clusters. In other words, all clients have the same label proportions, but clients in different clusters represent the same labels using different feature distributions. Formally, for any two clients di , dj ∈ Cq , holds: Pi (y) = Pj (y),
Pi (x | y) = Pj (x | y),
whereas for any two clients di ∈ Cq and dj ∈ Cr , with q ̸= r: Pi (y) = Pj (y),
Pi (x | y) ̸= Pj (x | y).
C. Random Network Distillation Random Network Distillation (RND) is a lightweight method for estimating the novelty of an input with respect to a set of previously observed examples [8]. This method is based on the observation that neural networks tend to produce lower prediction errors on inputs similar to those seen during training, and higher errors on unfamiliar or distributionally different inputs. Therefore, the prediction error of a network trained on past data can be used as a proxy for novelty, or equivalently as an estimate of uncertainty on previously unseen samples. Basically, RND consists of two neural networks. A target network, randomly initialized and kept fixed, defines a mapping (or embedding): f : X → Rk ,
(1)
while a predictor network fˆ(·; θ) : X → Rk
(2)
is trained to approximate the target network on the available data. Given a data distribution B, the predictor is optimized by minimizing the expected mean squared error 2 ∗ ˆ θ = arg min Ex∼B f (x; θ) − f (x) . (3) θ
2
Once trained, the RND uncertainty score of a sample x is computed as u(x) = fˆ(x; θ∗ ) − f (x)
2
. 2
(4)
Intuitively, the two network will produce similar outputs for inputs that are similar to the training data, while they will diverge on novel inputs. Thus, low values of u(x) indicate that x is similar to the data used to train the predictor, whereas high values suggest that x is novel or comes from a different distribution. In this work, we exploit this property to estimate similarity between clients in FL. A predictor trained on the local data of one client is expected to yield low uncertainty on data drawn from clients with similar feature distributions, and higher uncertainty on data from clients belonging to different clusters. RND therefore provides a compact and task-agnostic mechanism to compare local data distributions without exchanging raw data or evaluating the main federated model. III. M OTIVATION AND R EFERENCE S CENARIO Consider a set of clients collaboratively training an image recognition model from photos collected on mobile devices. Clients may share the same semantic task–recognizing cars, buildings, people, or street signs–while observing the world under substantially different visual conditions. Photos taken in different cities naturally differ in illumination, weather, architectural style, or sensor characteristics. As a result, clients
from the same city tend to benefit from collaborating with one another, whereas aggregating their updates with clients from visually different environments may produce a model that serves no group well. A symmetric setting arises in distributed sensing. Environmental monitoring devices deployed across heterogeneous areas may all measure the same physical phenomenon, yet under markedly different local conditions. Sensors placed near a coastal area, an urban district, or an industrial zone observe different noise patterns, dynamics, and baseline statistics. These scenarios expose two intertwined challenges for autonomous distributed systems, namely: (i) determining who should collaborate with whom; and (ii) allowing the corresponding federations to emerge without requiring their number to be specified externally. CFL addresses the first challenge by training specialized models for groups of compatible clients. However, existing approaches only partially address the autonomous discovery of such groups. In IFCA, the number of clusters is a hyperparameter that must be fixed before training; an incorrect choice may lead to suboptimal assignments and affect the quality of the final models—as shown in [7]. Moreover, cluster assignment is coupled with the primary learning task, since clients must evaluate full-size cluster models on their local data. PSFL relaxes the need to predefine a global number of clusters by allowing federations to emerge from local interactions, but still estimates client compatibility through the full task model. Thus, in both cases, discovering collaboration structures remains tied to repeated evaluations of the model being trained. In the scenarios above, this repeated coupling is largely unnecessary. The relevant client groups reflect relatively stable data-generation conditions, such as the physical environment in which sensors are deployed, that are unlikely to shift dramatically between rounds. Clustering can therefore be treated as a separate and cheaper preliminary step, performed once before federated training begins, or re-triggered occasionally when distributions change. The key challenge is then to estimate whether two clients observe compatible data distributions, without exchanging raw data and without relying on the primary model. Rather than asking whether a full task model trained by one client generalizes to another client’s data, we reframe the problem: does one client’s data appear novel with respect to another client’s local experience? IV. M ETHOD F ORMALIZATION A. Problem Formulation Let D = {di }N i=1 be the set of N devices participating in the FL process. Each device di holds a local dataset Bi : i Bi = {zij }m j=1 ,
zij = (xij , yij ) ∼ Pi .
(5)
where xij is the j-th input feature vector on device di , yij is the corresponding label, and mi is the number of local samples. Here, each local distribution Pi satisfies clustered feature skew, as defined in Definition 1.
The devices are partitioned into groups (clusters) C1 , . . . , Ck , with Cq ∩ Cr = ∅ for q ̸= r (no overlap) Sk and q=1 Cq = D (all devices belong to some cluster), however, this partition is not known to the devices at learning time. Formally, the goal is to learn a mapping M from the set of devices to a partition of D: M : D → F ⊆ P(D)
(6)
where P(D) denotes the power set of D, and M produces a set of federations F = {F1 , . . . , Ff }, where each federation Fq ⊆ D represents a group of devices that will collaboratively train a specialized model. The goal is that each discovered federation Fq matches as closely as possible the corresponding true cluster Cq . Formally, we aim to find M that minimizes the misclassification rate: f X 1 min Fq △Cπ(q) M N π∈Πk q=1
min
(7)
where Πk is the set of all permutations over {1, . . . , k}, △ denotes the symmetric difference between two sets, and | · | is the cardinality. The permutation π accounts for the fact that federation labels are arbitrary. Once M is determined, each federation Fq ∈ F independently runs a federated learning process, training a specialized model ωq optimized for the shared data distribution within that group. B. Novelty Driven Clustering This section details the proposed novelty driven federation discovery method, whose device-side and server-side procedures are summarized in Algorithms 1 and 2, respectively. Each device di trains a local RND predictor fˆi (·; θi ) by minimizing the expected prediction error with respect to a shared, randomly initialized target network f , as formalized in Equation (3). Once trained, each device broadcasts its predictor fˆi (·; θi∗ ) to all other devices. Each device di then evaluates every received predictor fˆj on its own local data Bi , obtaining a novelty score that quantifies the distributional divergence between dj and di : 2 sij = Ex∼Bi fˆj (x; θj∗ ) − f (x) , ∀j ∈ {1, . . . , N }. 2
(8) Low values of sij indicate that dj ’s predictor generalizes well to di ’s data, suggesting similar underlying distributions; high values indicate distributional divergence. Each device di then reports its row of scores {sij }N j=1 to the server, which reconstructs the full similarity matrix S ∈ RN ×N where S[i, j] = sij , without ever accessing raw local data. a) Adaptive Federation Discovery: the partition F can be derived directly from the observed novelty scores via a compatibility threshold. Each device di uses self-novelty sii as reference baseline, reflecting the residual prediction error on its own data after local training. Device dj is compatible with di if its cross-novelty score does not exceed the self-novelty by more than a relative margin: dj ∼ di ⇐⇒ sij ≤ sii + ϵ · σi ,
(9)
Algorithm 1 Device-Side RND Novelty Estimation Require: Device di with local dataset Bi Require: Shared random target network f Ensure: Local predictor fˆi and novelty row si 1: fˆi ← T RAIN RNDP REDICTOR(Bi , f ) 2: S END P REDICTOR (di , fˆi ) 3: {fˆj }N j=1 ← R ECEIVE P REDICTORS (di ) 4: for all fˆj ∈ {fˆ1 , . . . , fˆN } do 5: sij ← C OMPUTE N OVELTY(Bi , fˆj , f ) 6: end for 7: si ← (si1 , . . . , siN ) 8: S END N OVELTY ROW (di , si )
where σi is the standard deviation of {sij }N j=1 and ϵ ≥ 0 is a tolerance hyperparameter. The number of federations emerges naturally as the number of distinct compatibility groups identified across devices, rather than being specified in advance, and is expected to converge to the true number of clusters k. Once the partition F = {F1 , . . . , Ff } is determined, each federation Fq independently runs a federated learning process. b) Clustering Frequency: in stationary environments, the clustering phase needs to be performed only once prior to any federated training, since the data-generation conditions that define the cluster structure are not expected to change. When data distributions evolve over time [12], the clustering phase can be re-triggered periodically. We introduce a reclustering interval τ as a hyperparameter, expressed in number of federated rounds, such that the full RND-based clustering procedure is repeated every τ rounds. c) Decentralized Extension: while the formulation above assumes a central server that collects all predictors and computes the full similarity matrix S, the approach extends naturally to decentralized settings. A fully decentralized implementation can be realized by selecting the “server” via leader election, or by propagating the predictors across the network through a gossip protocol. Alternatively, each device di may exchange predictors only with its neighbors N (i), obtaining a partial view of the similarity matrix restricted to reachable peers. Cluster assignments can then be derived locally and reconciled into a globally consistent partition via a consensus protocol among neighboring devices, similarly to how decentralized aggregation has been approached in prior work [13], [14]. This decentralized variant preserves the core properties of the approach (i.e., no raw data exchange, decoupling from the main learning loop) while removing the assumption of a central coordinator. V. E XPERIMENTAL E VALUATION A. Experimental Setup The experimental evaluation is designed to validate whether the proposed novelty driven procedure can recover the latent groups induced by clustered feature skew. We used CIFAR-10 [15], a well-known benchmark dataset for computer vision tasks, and synthetically partitioned it
to reproduce the case study described in Section IV. The partitioning was performed using ProFed [16]. Given a number of groups k, the dataset was divided into k subsets while preserving the same label distribution in each group. Then, a different Gaussian noise distribution was applied to the images of each group, thereby inducing group-specific feature skew while keeping the underlying classification task unchanged. In our experiments, we set k = 4 and assigned three devices to each group, resulting in a total of N = 12 devices. For the RND predictors, we used a lightweight CNN composed of two convolutional layers followed by a flattening layer. After local training, each device evaluates both its own predictor and the predictors trained by all other devices on its local data, producing the novelty scores sij used to construct the novelty matrix. We then measured: (i) the resulting novelty scores, to assess whether devices within the same group exhibit similar residuals; and (ii) the total machine-time1 spent to perform clustering, including local RND training and inference with all received predictors. To contextualize the overhead of the proposed method, we also measured the clustering cost of IFCA. In this case, cluster assignment relies on evaluating the full task model rather than a lightweight auxiliary predictor. We used a ResNet18 [17] as the full classification model, since this architecture is commonly adopted for CIFAR-10 and is known to provide strong performance on this dataset. Each experiment was repeated with 10 different random seeds to account for stochasticity and avoid cherry-picking favourable runs. The reported results correspond to averages over these independent repetitions. The experimental code is publicly available2 .
Fig. 2. Difference between self-novelty sii and cross-novelty sij across devices. The colors of the device labels indicate the true clusters used to generate the experimental data. Near-zero values within diagonal blocks indicate that devices in the same group induce similar RND residuals. Cumulative time [s]
Algorithm 2 Server-Side RND-Based Federation Discovery Require: Devices D = {d1 , . . . , dN } Require: Compatibility tolerance ϵ ≥ 0 Require: Re-clustering interval τ Ensure: Discovered federations F 1: r ← 0 2: while training is not completed do 3: if r = 0 or r mod τ = 0 then 4: {fˆi }N i=1 ← C OLLECT P REDICTORS (D) 5: D ISTRIBUTE P REDICTORS(D, {fˆi }N i=1 ) ← C OLLECT N OVELTY R OWS (D) 6: {si }N i=1 7: S ← B UILD N OVELTY M ATRIX({si }N i=1 ) 8: F ← E XTRACT F EDERATIONS(S) 9: end if 10: RUN F EDERATED L EARNING(F) 11: r ←r+1 12: end while 13: return F
10 10 10
3
2
1
0
20
40
60
Global round
RND clustering RND every 10 rounds
80
100
RND every round IFCA
Fig. 3. Clustering overhead over federated rounds. Compared with IFCA, RND-based clustering reduces cost by re-clustering less frequently and by relying on small auxiliary predictors instead of full task models.
B. Discussion Figure 2 reports the results of the first experiment, where each device evaluates the RND predictors trained by all other devices on its own local data. In particular, the figure shows the difference between the local residual sii and the cross-novelty score sij . The diagonal block structure clearly highlights that, for devices belonging to the same group, this difference tends to be close to zero. This indicates that predictors trained within the same group produce similar residuals, and therefore generalize to each other’s data. Conversely, the difference remains higher across groups, showing that RND novelty scores capture the distributional divergence induced by feature skew. These results support the use of RND as a lightweight proxy for identifying which devices should collaborate. An important aspect of the proposed method is that the number of federations does not need to be specified a priori. 1 taken on an Intel Core i7 8700K, 64GB RAM, nVidia RTX 4070. 2 https://github.com/domm99/experiments-2026-uncertainty-basedclustered-fl
This is in contrast with approaches such as IFCA, where the number of clusters is typically fixed before training. In highly distributed scenarios, estimating this number can be difficult and an incorrect choice may affect the quality of the final models. Instead, in our approach, federations are derived from the structure of the novelty matrix through the compatibility criterion, allowing groups of similar devices to emerge from the observed data. Figure 3 reports the overhead associated with the clustering phase. The proposed RND-based approach exhibits substantially lower cost than IFCA, since clustering relies on compact auxiliary predictors rather than on the full task model used for FL. Moreover, RND-based clustering does not need to be executed at every global round: in stationary settings it can be performed once before training, while in non-stationary settings it can be repeated periodically. Even in the most conservative configuration, where RND clustering is repeated at every round, the overhead remains about one order of magnitude lower than IFCA in our experiments. A limitation of the centralized formulation is that each device evaluates the predictors received from all other devices. As the number of devices grows, this all-to-all evaluation may reduce the overhead gap with respect to full-model approaches. However, this is not intrinsic to the method. In large-scale distributed deployments, a decentralized implementation can restrict predictor exchange and evaluation to local neighborhoods. In that case, each device only evaluates the predictors of nearby peers, whose number is typically much smaller than the total population size. VI. C ONCLUSIONS AND F UTURE W ORK In this paper, we introduced an RND-based clustering mechanism for Clustered Federated Learning under structured feature skew. The proposed approach decouples federation discovery from the main learning process by relying on lightweight auxiliary predictors instead of full task models. As a result, devices can infer suitable collaboration groups from novelty estimates, without sharing raw data and without requiring the number of clusters to be fixed a priori. Preliminary experiments on synthetically partitioned CIFAR-10 show that RND residuals expose the latent clustered structure of the data and can separate devices affected by different feature perturbations. The results also indicate a substantially lower clustering overhead compared with IFCA, especially when clustering is performed once or periodically rather than at every global round. Future work will extend the evaluation to larger populations of devices, additional forms of non-IID data, and real-world distributed sensing scenarios. We also plan to investigate fully decentralized variants, where devices exchange RND predictors only with local neighbors. ACKNOWLEDGMENTS The authors used Generative AI tools for grammar and language editing assistance throughout the manuscript. All AIassisted text was reviewed and edited by the authors, who take full responsibility for the content of this paper.
R EFERENCES [1] B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-efficient learning of deep networks from decentralized data,” in Proceedings of the 20th International Conference on Artificial Intelligence and Statistics, AISTATS 2017, 20-22 April 2017, Fort Lauderdale, FL, USA, A. Singh and X. J. Zhu, Eds., vol. 54. PMLR, 2017, pp. 1273–1282. [2] Q. Li, Y. Diao, Q. Chen, and B. He, “Federated learning on non-iid data silos: An experimental study,” in 38th IEEE International Conference on Data Engineering, ICDE 2022, Kuala Lumpur, Malaysia, May 9-12, 2022. IEEE, 2022, pp. 965–978. [Online]. Available: https://doi.org/10.1109/ICDE53745.2022.00077 [3] P. Kairouz and et al., “Advances and open problems in federated learning,” Found. Trends Mach. Learn., vol. 14, no. 1-2, pp. 1–210, 2021. [Online]. Available: https://doi.org/10.1561/2200000083 [4] L. Esterle, “Deep learning in multiagent systems,” in Deep Learning for Robot Perception and Cognition. Elsevier, 2022, pp. 435–460. [5] E. Liu, W. Yang, Y. Gu, W. Long, S. István, and L. Jiang, “A survey of clustering federated learning in heterogeneous data scenarios,” Journal of Computing and Electronic Information Management, vol. 16, no. 3, pp. 17–22, 2025. [6] A. Ghosh, J. Chung, D. Yin, and K. Ramchandran, “An efficient framework for clustered federated learning,” IEEE Trans. Inf. Theory, vol. 68, no. 12, pp. 8076–8091, 2022. [Online]. Available: https://doi.org/10.1109/TIT.2022.3192506 [7] D. Domini, N. Farabegoli, G. Aguzzi, M. Viroli, and L. Esterle, “Decentralized proximity-aware clustering for collective self-federated learning,” Internet of Things, vol. 35, p. 101841, 2026. [Online]. Available: https://doi.org/10.1016/j.iot.2025.101841 [8] Y. Burda, H. Edwards, A. J. Storkey, and O. Klimov, “Exploration by random network distillation,” in 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net, 2019. [Online]. Available: https: //openreview.net/forum?id=H1lJJnR5Ym [9] A. Castagna and I. Dusparic, “Expert-free online transfer learning in multi-agent reinforcement learning,” in ECAI 2023 - 26th European Conference on Artificial Intelligence, September 30 - October 4, 2023, Kraków, Poland., K. Gal, A. Nowé, G. J. Nalepa, R. Fairstein, and R. Radulescu, Eds. IOS Press, 2023, pp. 357–364. [Online]. Available: https://doi.org/10.3233/FAIA230291 [10] H. B. McMahan, E. Moore, D. Ramage, and B. A. y Arcas, “Federated learning of deep networks using model averaging,” CoRR, vol. abs/1602.05629, 2016. [Online]. Available: http://arxiv.org/abs/ 1602.05629 [11] A. Nilsson, S. Smith, G. Ulm, E. Gustavsson, and M. Jirstrand, “A performance evaluation of federated learning algorithms,” in Proceedings of the Second Workshop on Distributed Infrastructures for Deep Learning, DIDL@Middleware 2018, Rennes, France, December 10, 2018. ACM, 2018, pp. 1–8. [Online]. Available: https://doi.org/10.1145/3286490.3286559 [12] D. Domini, G. Aguzzi, L. Pellegrini, M. Viroli, and L. Esterle, “C2fl: Clustered continual federated learning under spatial and temporal drift,” CoRR, vol. abs/2606.18003, 2026. [Online]. Available: https://doi.org/10.48550/arXiv.2606.18003 [13] D. Domini, G. Aguzzi, L. Esterle, and M. Viroli, “FBFL: A field-based coordination approach for data heterogeneity in federated learning,” Log. Methods Comput. Sci., vol. 22, no. 1, 2026. [Online]. Available: https://doi.org/10.46298/lmcs-22(1:19)2026 [14] D. Domini, G. Aguzzi, A. D. Zenoozi, L. Erhan, L. Cavallaro, A. Liotta, and M. Viroli, “Sparseful: Self-organizing sparse federated learning over spatially non-iid data.” SSRN, 2026. [Online]. Available: https://ssrn.com/abstract=6584516 [15] A. Krizhevsky, G. Hinton et al., “Learning multiple layers of features from tiny images,” 2009. [16] D. Domini, C. O. Ingemann, G. Aguzzi, L. Esterle, and M. Viroli, “Profed: a benchmark for proximity-based non-iid federated learning,” Joural of Open Research Software, vol. 14, 2026. [Online]. Available: https://openresearchsoftware.metajnl.com/articles/10.5334/jors.624 [17] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016. IEEE Computer Society, 2016, pp. 770–778. [Online]. Available: https://doi.org/10.1109/CVPR.2016.90