Distributed Rumor Source Detection via Boosted Federated Learning

How to localize the rumor source is an extremely important matter for all sectors of the society. Many researchers have tried to use deep-learning-based graph models to detect rumor sources, but they have neglected how to train their deep-learning-based graph models in the noisy social network environment efficiently. Especially for deep learning models, the performance relies on the data scale. However, even though it is known that a substantial amount of rumor data distributed across multiple edge servers (e.g., cross-platform), due to conflicting business interests, its challenging to coordinate all parties to train a model driven by many samples while avoiding moving data. Federated learning, is an effective technique to bridge this gap. Therefore, this paper proposes a Distributed Rumor Source Detection via Boosted Federated Learning (DRSDBFL). Specifically, this paper proposes an effective rumor source detection method based on a deep-learning-based graph model with a denoising module. To the best of our knowledge, we are the first to attempt the use of a denoising module to reduce the noisy effects of social networks. Then, we propose a novel boosted federated learning mechanism through boosting the high-quality edge worker to improve the training efficiency. Finally, the effectiveness of the proposed method is verified by extensive experiments.

I. Introduction

In the past few years, with the popularity of mobile devices, the number of social media users represented by Twitter, Facebook and Weibo has surged [1]. Due to the wide coverage of users and the variety of content spread by users, this brings a huge burden to the supervision of public opinion on social networks; furthermore, it has become very difficult to locate the source of rumors. Fortunately, immediately after Wang et al. [2] proposed using a label propagation algorithm to simulate the reverse propagation process of rumors without specifying the underlying propagation model (LPSI), a large number of rumor source detection methods based on graph models have emerged [3], [4], [5]. In the face of the current complex social network environment, it is very promising to use the deep-learning-based graph model, a technology with strong generalization ability, to realize a data-driven rumor source detection method. However, there are still some problems, as follows: 1)

Uncertainty: Previous works usually utilize epidemic models, assuming that the diffusion pattern of rumors in social networks is known. However, it is difficult to predict the diffusion mode of rumors in social networks with strong uncertainties in advance in the real world [5].

2)

Noise: Most of the existing methods focus on how to simulate the spread of rumors in social networks, but like LPSI and GCNSI methods, they ignore the noisy nature of social networks [6]. On social platforms, as users are flooded with content on various themes, the themes of this content are most likely not related to the current rumors for which the source needs to be traced. When considering the diffusion of the current theme event, other theme events essentially act as noise [7], interfering with our line of sight. How to eliminate this noise to provide higher quality for the final service is worth studying.

3)

Training efficiency: The social network data is huge, even for the same group of people due to the diversity of social network events or the platform [8], the scale of related event spreading data is quite large. The recently emerging federated graph learning technology only optimizes training from the perspective of node-level, which improves training efficiency but loses some edge information in the global graph [9]. How to divide graph data from other perspectives like graph-level and improve model training efficiency is a new challenge in the future.