Less is More: A privacy-respecting Android malware classifier using federated learning

Abstract In this paper we present LiM (‘Less is More’), a malware classification framework that leverages Federated Learning to detect and classify malicious apps in privacy-respecting manner. Information about newly installed kept locally on users’ devices, so the provider cannot infer which were by users. At same time, input from all users taken into account federated learning process they benefit better performance. A key challenge of setting do not have access ground truth (i.e. correctly identify whether an app malicious). To tackle this, uses safe semi-supervised ensemble maximizes accuracy with respect baseline classifier trained service cloud). We implement show cloud server has F1 score 95%, while clients perfect recall only 1 false positive > 100 apps, using dataset 25K clean 200 50 rounds federation. Furthermore, conduct security analysis demonstrate robust against both poisoning attacks adversaries who control half clients, inference performed honest-but-curious server. Further experiments Ma-MaDroid’s confirm resistance performance improvement due