Towards privacy-preserving cybersafety tools by using Federated Learning

Abstract

Living in the digital era, people can access a huge amount of online content daily. Online services might bene t humanity by easing many everyday life tasks. However, people and especially minor users can en- counter many threats while they are online. Despite various cybersafety tools and applications, the number of minors experiencing online threats is not decreasing. This work focuses on cybersafety tools that use Ma- chine Learning algorithms for automatic detection of inappropriate con- tent. Such tools require the collection of big of data that are often sen- sitive. Additionally, keeping these datasets up-to-date and retraining the models can be challenging. We propose using Federated Learning (FL) training to overcome these challenges. FL allows training a model on dis- tributed data without transferring data to a central unit. We provide a conceptual mapping between the components of a cybersafety framework architecture and the actors in the FL communication protocol to explain how FL can be applied in the context of cybersafety tools. We design and implement a TensorFlow-Federated simulation to explore FL training on a text classi cation model that detects aggressive text. We experimented with a centralized dataset of aggressive tweet posts to assess the perfor- mance of the model trained in the FL approach compared to a model trained in the centralized approach and explore how the number of clients participating in FL a ects the model's performance. Additionally, we ex- perimented with a local model training to assess the client device's CPU utilization, memory consumption, and execution time. The results show that the model's performance when trained in FL settings can approach the model's performance when trained in the traditional approach. Also, the model's performance improves when more clients participate in the FL training. Regarding the performance of a client's device, the results show that the execution time for a local model's training is short and does not over-consume the device resources. The ndings show that cybersafety tools are an applicable use case for FL training.