Detecting and responding to hostile disinformation activities on social media using machine learning and deep neural networks

Abstract

Disinformation attacks that make use of social media platforms, e.g., the attacks orchestrated by the Russian “Internet Research Agency” during the 2016 U.S. Presidential election campaign and the 2016 Brexit referendum in the UK, have led to increasing demands from governmental agencies for AI tools that are capable of identifying such attacks in their earliest stages, rather than responding to them in retrospect. This research undertaken on behalf of the Canadian Armed Forces and Department of National Defence. Our ultimate objective is the development of an integrated set of machine-learning algorithms which will mobilize artificial intelligence to identify hostile disinformation activities in “near-real-time.” Employing The Dark Crawler, the Posit Toolkit, TensorFlow (Deep Neural Networks), plus the Random Forest classifier and short-text classification programs known as LibShortText and LibLinear, we have analysed a wide sample of social media posts that exemplify the “fake news” that was disseminated by Russia’s Internet Research Agency, comparing them to “real news” posts in order to develop an automated means of classification.

Introduction to the Minitrack on Emerging Topics in Digital Government

“…LibLinear (LBL), the companion software for LST, attained a weighted average F1 of 0.827 for the Reddit dataset (Table 5), a noticeable improvement upon NB and SVM, but still below that attained when the Posit results were input to J48 (F1 = 0.863) and RF (F1 = 0.931). LST, on the other hand, continues to be an exceptionally solid performer when it comes to the automated machine classification of social media posts (Yang, 2017), as reported elsewhere in other dis/misinformation studies conducted by the ICCRC and Strathclyde (Cartwright et al, 2019;Cartwright et al, 2022). In this instance, with the Reddit dataset, LST attained a weighted average F1 of 0.982 (Table 6), better than Posit at 0.931, and much better than the weighted averages of the other machine-learning algorithms, which ranged from a low of 0.64 to a high of 0.827.…”

Information Trolls and Democracy: A Qualitative Examination of Disinformation Campaigns in Canada

“…While algorithms are not all bad, as they can also be employed to target and take down disinformation online (Cartwright et al, 2022;Sun, 2023), the power they possess to control the information environment is often exploited by information trolls and social media capitalism (Sun, 2023). The exploitation of algorithmic tools to spread disinformation coupled with the corporate goals of social media networks to maximize profits illustrates the online information environment's immense power in influencing public opinion.…”