Poligraph: Intrusion-Tolerant and Distributed Fake News Detection System
Guohou Shan,Boxin Zhao,James R. Clavin,Haibin Zhang,Sisi Duan
DOI: https://doi.org/10.1109/tifs.2021.3131026
IF: 7.231
2022-01-01
IEEE Transactions on Information Forensics and Security
Abstract:We present Poligraph, an intrusion-tolerant and decentralized fake news detection system. Poligraph aims to address architectural, system, technical, and social challenges of building a practical, long-term fake news detection platform. We first conduct a case study for fake news detection at authors’ institute, showing that machine learning-based reviews are less accurate but timely, while human reviews, in particular, experts reviews, are more accurate but time-consuming. This justifies the need for combining both approaches. At the core of Poligraph is two-layer consensus allowing seamlessly combining machine learning techniques and human expert determination. We construct the two-layer consensus using Byzantine fault-tolerant (BFT) and asynchronous threshold common coin protocols. We prove the correctness of our system in terms of conventional definitions of security in distributed systems (agreement, total order, and liveness) as well as new review validity (capturing the accuracy of news reviews). We also provide theoretical foundations on parameter selection for our system. We implement Poligraph and evaluate its performance on Amazon EC2 using a variety of news from online publications and social media. We demonstrate Poligraph achieves throughput of more than 5,000 transactions per second and latency as low as 0.05 second. The throughput of Poligraph is only marginally (${4\%}$ –${7\%}$ ) slower than that of an unreplicated, single-server implementation. In addition, we conduct a real-world case study for the review of fake and real news among both experts and non-experts, which validates the practicality of our approach.
computer science, theory & methods,engineering, electrical & electronic