Blog

Federated learning is a promising concept for owners of machine-learning models and owners of training data alike. Current federated learning approaches require the model owner to have knowledge of, and access to, the network of training devices. This requires him to create persistent connections to the devices and develop automated procedures for generating the data, while providing no incentive for data owners to participate. We outline a framework for orchestrating federated learning and rewarding data owners that does not rely on trust or knowledge between the model owner and data owners.

Adversarial learning is a new and growing area of machine-learning research. Formulating it using tools from game theory allows for a different view of machine learning, when compared to the traditional, purely statistical view. This view allows us to extend the scope of machine learning to security systems and privacy protection through the study of adversarial attacks, which allows for dealing with violations of the i.i.d assumption of machine learning that occur in some contexts. However, adversarial learning also raises some questions about the state and the future of machine learning. This survey provides a look at the intersection of machine learning and game theory by introducing the latter, then showing its application in adversarial learning. The survey continues with an overview of adversarial attacks and their known defenses, as well as prevention mechanisms for cases where sensitive data must not run the risk of being leaked by an adversarial attack.