PREDISTINE: A Privacy-Preserving and Resilient Framework for Distributed Modular Neural Networks on the Tiny Edge

Due to the large amount and diversity of data generated by Internet of Things (IoT) devices, Artificial Intelligence (AI) has seen a breakthrough in IoT applications in recent years. Traditionally this implies that data must be sent to the cloud, which increases communication costs, causes delays in system response and makes data vulnerable to privacy breaches. A solution is to migrate AI models to constrained devices. However, these devices often have limited computational resources and have a limited ability to implement AI models. Besides that, the system also needs to be flexible when devices fail or reconnect. Therefore, this research project aims to develop a novel framework for distributing neural networks and enhancing their resilience on constrained edge devices. During the realization of the framework, research will be done on how a model can be efficiently distributed over a network of edge devices, considering energy consumption, memory footprint and bandwidth usage. Secondly, the project will explore ways to make neural networks adaptable to various combinations of input devices, making the inference more resilient in case of failing devices. Lastly, the integration of distributed learning will be investigated to enhance model performance and personalize predictions without sacrificing privacy. To demonstrate the valorization potential of this research, all results will be implemented and validated on two real-world use cases, from which a software framework will be developed and made publicly available for use in other applications.