Anomaly detection at the edge: a federated learning approach

Details

Supervisors

Riviere, Etienne ; Klein, Sarah

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] en sciences informatiques, à finalité spécialisée

Abstract

In recent years, the development of many IoT applications has risen steadily, leading to many new horizons and possibilities. However, this growth also gave birth to new concerns, among these are privacy concerns [ Khan and Salah, 2018 ]. While ensuring secure and private communication can protect devices from the outside of their networks, some applications such as machine learning still need some form of model centralization and thus puts the data at risk of leaking to actors within the IoT networks. Fortunately, an emerging paradigm, known as federated learning aims to remove the data leak risk from decentralized machine learning [ Yang et al., 2019 ]. Its aim is to make clustering, regression, classification and other machine learning possible without sharing data. This master’s thesis explores the task of learning to detect anomalies on a network in a privacy preserving manner. In more details, we intend to tackle the problem of live anomaly detection on a connected fleet of IoT devices’ data distribution. The goal is to send an alert to the federated learning network when a new data point from a device does not fit the currently estimated global distribution. Due to the live nature of the problem, we cannot just ignore the temporal factor of the data. More specifically, we are going to use time-series to represent our data and use time series clustering techniques in a federated context. As neural networks tend to lack interpretability, we are going to use more interpretable clustering techniques such as KMeans [ Lloyd, 1982 ] and Agglomerative clustering [Aghabozorgi et al., 2015]