Federated learning (FL) is a promising paradigm that is gaining traction in the field of privacy-preserving machine learning for edge computing systems. With FA, multiple data owners, known as clients (for example, organizations in an inter-silo FA setting), can collaboratively train a model on their private data without having to transmit their raw data to external service providers. FA has been rapidly adopted in several thriving applications such as digital health, which generates the largest volume of data in the world. Decentralized learning (DL) goes a step further by offering serverless federated learning, where data is retained at the clients and no server is required. Thus, DL involves distributed and decentralized protocols to enable clients to build a global model. Although AD represents a first step toward protecting privacy by storing data locally on each client’s device, this remains insufficient because the model parameters shared by AD are vulnerable to privacy attacks, as demonstrated by recent research [8]. Furthermore, deep learning is more vulnerable to malicious behavior by clients who may inject corrupted information into the data and models, resulting in dysfunctional and unreliable deep learning models. Recent studies show that robustness and privacy in deep learning can be trade-offs; treating them independently, as is generally the case, can have reciprocal negative side effects.
Consequently, a new multi-objective approach is needed to ensure the robustness of data-flow models and protect them against privacy breaches. This project addresses this challenge and aims to specifically tackle the issues arising at the intersection of privacy, robustness, and the utility of deep learning models, through: (i) new deep learning protocols; (ii) a multi-objective approach that balances privacy, robustness, and utility, as these objectives are often conflicting; (iii) the application of these techniques to deep learning in continuous edge-to-cloud systems.