Detecting malicious clients in federated learning

Dominik Kolasa

supervisor: Maria Ganzha

Federated learning is a process where many clients contribute to the global model. In this scenario I found that some of the clients may intentionally or by a mistake provide wrong updates to the global model. I address this issue to filter such contributions. I am trying to answer the question of how to compare the models and provide only useful updates. In developed a new method where no testing data is required. I can assure security and minimize any mistakes with the dataset itself. For wrongly labeled or bad quality datasets the client’s model will differ from all the other models trained using proper data. My algorithm can detect such clients and do not include their contribution to the global model. The global model will learn the most common pattern along all the clients. The research shows that the method works for different setups, filters out malicious clients accurately when compared with other methods and does not require any data to run.

Federated learning is a new way of distributed machine learn-ing multiple clients. It can be used to train models using mo-bile phones, IOT devices and any device which has the power to compute a training round. All of them collaborate to build a global model. In my research I am trying to resolve problems with malicious updates which may affect federated learning process and cause the global model to predict even randomly.