Deep learning pervades heavy data-driven disciplines in research and development. The Internet of Things and sensor systems, which enable smart environments and services, are settings where deep learning can provide invaluable utility. However, the data in these systems are very often directly or indirectly related to people, which raises privacy concerns. Federated learning (FL) mitigates some of these concerns and empowers deep learning in sensor-driven environments by enabling multiple entities to collaboratively train a machine learning model without sharing their data. Nevertheless, a number of works in the literature propose attacks that can manipulate the model and disclose information about the training data in FL. As a result, there has been a growing belief that FL is highly vulnerable to severe attacks. Although these attacks do indeed highlight security and privacy risks in FL, some of them may not be as effective in production deployment because they are feasible only given special—sometimes impractical—assumptions. In this paper, we investigate this issue by conducting a quantitative analysis of the attacks against FL and their evaluation settings in 48 papers. This analysis is the first of its kind to reveal several research gaps with regard to the types and architectures of target models. Additionally, the quantitative analysis allows us to highlight unrealistic assumptions in some attacks related to the hyper-parameters of the model and data distribution. Furthermore, we identify fallacies in the evaluation of attacks which raise questions about the generalizability of the conclusions. As a remedy, we propose a set of recommendations to promote adequate evaluations.

Federated learning (FL) enables a set of entities to collaboratively train a machine learning model without sharing their sensitive data, thus, mitigating some privacy concerns. However, an increasing number of works in the literature propose attacks that can manipulate the model and disclose information about the training data in FL. As a result, there has been a growing belief in the research community that FL is highly vulnerable to a variety of severe attacks. Although these attacks do indeed highlight security and privacy risks in FL, some of them may not be as effective in production deployment because they are feasible only under special-sometimes impractical-assumptions. Furthermore, some attacks are evaluated under limited setups that may not match real-world scenarios. In this paper, we investigate this issue by conducting a systematic mapping study of attacks against FL, covering 48 relevant papers from 2016 to the third quarter of 2021. On the basis of this study, we provide a quantitative analysis of the proposed attacks and their evaluation settings. This analysis reveals several research gaps with regard to the type of target ML models and their architectures. Additionally, we highlight unrealistic assumptions in the problem settings of some attacks, related to the hyper-parameters of the ML model and data distribution among clients. Furthermore, we identify and discuss several fallacies in the evaluation of attacks, which open up questions on the generalizability of the conclusions. As a remedy, we propose a set of recommendations to avoid these fallacies and to promote adequate evaluations.