Lingyu Qiu (UNINA) presented his research on one-shot federated learning (OFL), focusing on developing efficient and trustworthy decentralised machine-learning methods that reduce communication overhead while addressing substantial differences between data held by individual participants. His work contributes to WP4 by investigating how federated models can be trained without centralising local data and without requiring repeated synchronisation between clients and a central server. The research is particularly relevant to distributed and industrial IoT environments, where communication bandwidth, data heterogeneity, privacy and computational resources can limit the effectiveness of conventional federated learning.

He presented a research programme comprising five connected contributions, progressing from a systematic review of 94 studies on one-shot federated learning to methods for data synthesis, pretrained-model integration, feature alignment and spatiotemporal prediction. The first contribution established a taxonomy of existing OFL approaches and identified pretrained models and decentralised generative synthesis as promising directions. FedSDE uses local conditional diffusion models to generate challenging samples and combines them with self-distillation to improve learning under highly non-IID data distributions. Subsequent work investigated the role of pretrained vision transformers, identifying a mismatch between their feature representations and conventional server-side aggregation. To address this problem, FedOPAL applies lightweight visual prompt tuning to align client-side representations while retaining closed-form aggregation and avoiding iterative server-side training. The final contribution, SWIFT, extends the approach to spatiotemporal traffic prediction by separating city-specific spatial information from temporal patterns that can be shared across participants.

The presentation reported results across image-classification and traffic-prediction benchmarks. For FedSDE, the reported absolute accuracy improvement over ensemble baselines was 16.57 percentage points on CIFAR-100 under the specified non-IID setting. FedOPAL achieved 93.72% accuracy on CIFAR-10, with no server-side training iterations, while SWIFT reduced communication bandwidth by a factor of 300 compared with iterative multi-round federated learning in the reported traffic-prediction experiments. The work has resulted in five scientific outputs, reported as published, including a systematic review in Neurocomputing, studies presented at IEEE BigData 2025 and ACM MobiCom 2025, work presented at FLICS 2026, and a contribution on spatiotemporal traffic prediction in Computer Networks. Lingyu also discussed limitations and future research, including the computational cost of local diffusion training, generalisation across domains, changing network topologies and the need to introduce formally quantified differential-privacy guarantees. These directions underline the importance of balancing predictive performance, efficiency and privacy in practical federated-learning systems.

Towards_Trustworthy_and_Efficient_One-Shot_Federated_Learning.pdf

S. M. Asiful Huda (UNINA) presented his research on decentralised graph neural networks (GNNs), focusing on federated training, model adaptation and interpretability without centralising data from individual participants. His work contributes to WP4 by addressing the challenges of learning from distributed graph-structured data, particularly when data are heterogeneous across clients and the relationships between nodes cannot be fully shared. The presentation focused on two complementary research directions: identifying and removing the influence of selected training data from federated GNNs, and developing shared structural concepts to improve the interpretability and cross-domain adaptability of graph foundation models.

The first direction introduced GFGU (Gated Federated Graph Unlearning), a framework designed to remove the influence of selected clients or nodes without retraining the entire model. Motivated by the right to erasure under Article 17 of the GDPR, the approach uses learnable gates after the first graph convolutional layer to isolate client- and node-level contributions. The evaluation covered several graph datasets, including citation networks, co-authorship graphs, e-commerce networks and crowdsourcing data. The reported results showed that GFGU achieved the highest node-classification accuracy among the compared methods on five of seven datasets for client-level unlearning and three of seven datasets for node-level unlearning. The evaluation also considered forgetting quality, similarity to a fully retrained model and vulnerability to membership-inference attacks.

The second direction introduced FedConcept, a framework for developing federated graph foundation models using a shared vocabulary of structural concepts. It represents graph structure through degree percentiles, which capture the relative connectivity of nodes within their respective domains, and uses these concepts as an interpretable intermediate representation for federated learning and downstream adaptation. Preliminary experiments covered citation networks, Wikipedia graphs, knowledge graphs and molecular datasets. FedConcept achieved the highest reported results on several benchmarks, including Cora, WikiCS, FB15k-237, WN18RR, HIV and PCBA, although performance varied across datasets. The presentation also reported a published survey on graph-based federated learning for smart healthcare, a submitted paper on architecture-aware unlearning in federated graphs, and ongoing work on graph foundation models. Overall, the research demonstrated two implemented and evaluated approaches that seek to make the influence of decentralised graph models identifiable and interpretable, while identifying uncertainty and interpretability in federated GNNs as important topics for further investigation.

Decentralised_Graph_Neural_Networks_AdaptationTraining_and_Interpretability_in_Federated_Environments.pdf