Leader: Francesco Piccialli (UNINA)

Objectives - Within the overarching goal of achieving a deeper understanding of artificial intelligence through transparent, open and explainable perspectives, the aim of this WP will be to innovate, investigate and implement advanced Federated Learning methodologies in generative contexts and to investigate the evolution and advancements of Graph Neural Networks (GNNs).  The three-fold aim is:

• To advance the understanding and development of the federated learning mechanisms that support privacy and security, while simultaneously facilitating the generation of novel models and data.
• To augment the capacity of GNNs to process graph-structured information, which will make these networks more transparent and efficient.
• To integrate NLP in Federated Settings: to explore the integration of NLP with federated learning and GNNs, enhancing the ability to process and generate natural language data in a decentralized environment.

Tasks:

T1 Development and Exploration of Federated Learning for Generative Models

This task investigates the integration of federated learning with generative models such as Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs). The primary focus is to design and refine  architectures that effectively combine the data generation power of generative models with the decentralised, privacy-preserving nature of federated learning. The process begins with the development of prototype architectures that are able to adapt to the unique requirements of decentralised data sources. A significant emphasis is placed on ensuring that these federative generative models maintain the quality of data generation while upholding stringent privacy and security standards. This includes identifying and mitigating the potential privacy risks and security vulnerabilities that are unique to federated generative scenarios. Another critical aspect of this task is ensuring the efficiency and scalability of the models. Given the inherent challenges of managing data across multiple decentralised nodes, it is crucial to develop systems that are not only robust in terms of data processing and generation but that are also able to scale effectively across different network architectures and data volumes. This task requires a thorough understanding of both federated learning and generative modelling along with a strong emphasis on privacy, security, efficiency and scalability.

T2 Advanced Studies on Graph Neural Networks (GNNs) with Federated Learning

In this task, the focus will be on the advanced study and adaptation of Graph Neural Networks (GNNs) within the federated learning paradigm. The goal is to enable GNNs to efficiently process and analyse graph-structured data in a decentralised manner. This task will involve adapting and modifying GNN architectures to ensure that they are suitable for federated settings, where data is not centrally stored but is distributed across various nodes. The key challenges include maintaining the integrity and effectiveness of the GNNs while they operate on partitioned data and addressing issues related to model synchronisation and efficient data communication in decentralised environments. Additionally, improving the interpretability of GNNs in these settings is also a major focus. Given the complexity of graph-structured data and the intricacies of GNN operations, it is crucial to develop methods that can provide clear insights into how these networks process and interpret data. This will involve exploring new approaches to model transparency and explainability, particularly in the context of decentralised data structures.

T3 Integrating NLP with Federated Learning and GNNs

This task is centred around the innovative integration of Natural Language Processing (NLP) with federated learning frameworks and Graph Neural Networks (GNNs). The challenge here will be to develop NLP models that can effectively function in a federated learning environment, thereby leveraging the structural data processing capabilities of GNNs. This will involve designing NLP algorithms and systems that are able to handle the complexities of language processing and generation while operating across decentralised data sources. The task comprises tackling various challenges, including efficiently distributing the linguistic data across nodes, ensuring the privacy and security of the data during processing and maintaining the performance of NLP models in a non-centralised setup. Another important aspect will be to improve the interpretability of these NLP models in federated environments. This would require innovative approaches to explain how these models process natural language data and make decisions, thereby making the NLP processes more transparent and understandable in a decentralised context.

T4 Real-world Implementation and Testing of Developed Frameworks

The final task will involve the practical application and real-world testing of the frameworks and models that were developed in the previous tasks. This will include implementing the federated generative models, advanced GNNs and integrated NLP systems in actual, operational environments. The key focus will be on moving beyond the theoretical models and controlled environments in order to deploy these systems in real-world scenarios, potentially in collaboration with various industries and sectors. This task is crucial for evaluating the practical viability and effectiveness of the developed models. It will involve establishing rigorous criteria for performance assessment, including the metrics for accuracy, efficiency, privacy preservation and overall system robustness. An important component of this task will be to establish feedback loops from real-world applications, which are essential for the iterative refinement and improvement of the models and methodologies. This stage is critical for translating theoretical research into practical, scalable and efficient AI solutions that can be widely adopted across different domains.