Summaries - Office of Research & Innovation
Research Summaries
Back Artificial Neural Networks for Automated Detection of Hostile Information Campaigns
Fiscal Year | 2019 |
Division | Graduate School of Operational & Information Sciences |
Department | Defense Analysis |
Investigator(s) | Warren, Timothy C. |
Sponsor | Defense Intelligence Agency (DoD) |
Summary | New approaches to machine learning now make it possible to detect, assess, and predict rapidly evolving hostile information campaigns on a global basis, by training artificial neural networks to detect anomalous patterns in human communications that precede episodes of radicalized violence. Current approaches to online media exploitation are largely anecdotal, with most products from the private market offering no systematic means of validating their metrics, because the underlying data and algorithms arc held on a proprietary basis. In contrast, we have proposed development of a platform which integrates diverse open-source data streams to generate global-scale predictions of emergent threats to human security. By generating systematic and transparent predictions of the timing and location of collective violent events, this approach provides metrics of radicalization that are both transparent and directly verifiable. Our research aims to provide new tools toots allowing us to detect rapidly evolving hostile efforts at manipulation, and to predict the emergence of sites of unrest and contestation prior to the outbreak of violence. To do so, our approach conducts cross-lingual analysis of publicly available social media and mass media sources, to generate systematic spatio-temporal maps of social and political radicalization in human discourse on a global basis. These maps then provide the inputs to our second stage of analysis, which utilizes artificial neural network algorithms to generate adaptive sensors which learn to detect anomalous patterns as early indications of hostile information campaigns and emergent threats to human security. In particular, we have proposed novel machine learning algorithms, including hybrid applications of convolutional neural networks and quasi-recurrent neural networks, to create a new generation of contextually aware and culturally nuanced "sensors" that respond dynamically to unforeseen anomalies in incoming data streams. Our platform trains neural network models that can automatically learn the normal patterns of discourse in a spatial region, even as communication patterns dynamically evolve. By detecting sudden, radicalized deviations from these normal communication patterns, such models can then be used to identify early warning signs of unrest and fear on complex human landscapes. Moreover, because such models can learn updated patterns as new forms of rhetoric and slang emerge, they can provide warning signals even i n regions where human knowledge of recent shifts in local dialects may be lacking. |
Keywords | Information Operations machine learning |
Publications | Publications, theses (not shown) and data repositories will be added to the portal record when information is available in FAIRS and brought back to the portal |
Data | Publications, theses (not shown) and data repositories will be added to the portal record when information is available in FAIRS and brought back to the portal |