Collaborative (federated) machine learning has recently emerged as a promising approach for building machine learning models using distributed training data held by many parties. In this setting, the training algorithm is also distributed, and participants repeatedly exchange information about their data in the context of the AI task, through some aggregator servers. The objective of such an algorithm is to enable all participants converge to a global model, while their data remain local. Thus, this approach is very attractive to parties that own sensitive data, and agree on performing a common AI task, yet are unwilling to pull their data together for centralized training of a model.
There are severe obstacles limiting the widespread deployment of secure, efficient, and truly privacy-preserving collaborative AI. Recent research results show that the existing collaborative AI algorithms can leak a significant amount of sensitive information about local datasets, are not robust to noisy and heterogeneous data, are susceptible to adversarial interventions, and impose significant communication and computation costs on the participants.
In this project, we propose Efficient and Secure Collaborative Artificial Intelligence (ESCAI), a framework for large-scale distributed machine learning with low computation and communication overhead, provable data privacy, and strong robustness guarantees against adversarial entities.