Terrorist Finger Prints: Using AI to Identify Terrorist Group Perpetrators

• In Event: Terrorism

Abstract

Terrorism represents an undesirable but seemingly inevitable part of the modern social landscape. An examination of terrorist dynamics can provide useful insights into counter-terrorism policies that are both more effective at reducing violence and less invasive on society. With the tremendous increases taking place in Artificial Intelligence (AI) capabilities, innovative approaches to analyzing terrorist data can yield valuable information regarding the interaction of terrorists, governance and society.

Over half of terrorist attacks globally have no known perpetrator (GTD 2018). The inability to identity the perpetrator of a terrorist action restricts counterterrorism operations, allows terrorist groups to ‘fly under the radar,’ and creates on its own a level of fear resulting from the terrorist attack. It is crucial that we develop better means to identify the perpetrators of terrorist violence. We apply sophisticated AI approaches to intrastate terrorism data integrated with extensive demographic, ethnic, population, economic, religious, geographic and political information to answer: To what extent can the array of descriptors of an attack, its location and timing enable identification of the likely terrorist group?

Four factors have limited the application of AI approaches to terrorism. First, quantitative studies of terrorism primarily conduct cross-national studies using country based annual data (Enders et al. 2011). As a result, they tend to have few observations, limiting the applicability of data intensive AI approaches. Second, terrorism analyses tend to analyze a single database and rarely merge those data with other information. Third, scholars from areas where sample size supported neural networks and other deep approaches, have unsuccessfully applied their models to the more limited data sets of terrorism violence. Finally, a more comprehensive approach is by nature transdisciplinary and requires a coordinated, interdisciplinary team. We address each these issues.

First, employing disaggregated, country-level data over time on states that experience high levels of terrorism provides us with a large and highly-varied set of observations typically not studied. Countries such as India, Nigeria, and Pakistan have experienced thousands of terrorist attacks over the last decade providing an extensive array of incident data (Piazza 2009). We propose to build a global database and understanding of terrorism by knitting together a series of large-N, country-specific data sets that provide both more and more detailed information on dynamics.

Second, we augment each country-specific terrorism datasets with extensive arrays of information on country attributes. Merging terrorism data with country-specific census and other datasets, we capture critical social, political, economic and demographic dynamics (Wilson and Piazza 2013). These combined data allow us to embed our study of terrorism within the detailed context of a country while still employing sophisticated statistical approaches.

Third, we employ AI approaches that appropriately match our data parameters (Bowie et al. 2017). We opted initially to apply initially shallow machine learning models such as Decision Trees and Random Forest (Biau 2016), to build models that require only a modest amount of data to train themselves. Our machine learning models will use the open source sci-kit learn library, scikit-learn - Machine Learning in Python (http://scikit-learn.org/) and be used to predict likely targets and perpetrator groups. With these Machine Learning approaches we examine the dynamics within the matrixes of our combined datasets, revealing hard to see but critical patterns.

Finally, we have put together a transdisciplinary team. Led by scholars with different disciplinary backgrounds and drawing on a group of highly skilled and varied experts.

We have begun to employ this approach successfully. In a published pilot study, we were able to integrate the Global Terrorism Database (GTD) with demographic information from the Census in India (which has data on all India’s 640 districts) and apply Decision Tree and Random Forest AI to questions about terrorist event location. The result was models that were dramatically more accurate in predicting where terrorist events would occur than previous models (Verma et al. 2019). Factors such as location, religion, literacy and population size provided substantial explanatory power for predicting terrorism location.

We next turn our analyses to the question of how to identify terrorist groups by their AI-data signatures (McMillan et al. 2019). Using merged datasets, and Decision Tree and Random Forest artificial intelligence models initially on India and Nigeria, we will try to identify ‘AI data-signatures’ for actors conducting terrorist attacks. These signatures would examine the characteristics of the attacks, their location, type, deadliness, other attacks conduct by the group and additional factors that can inform our models predicting identify.

Author

• Scott Sigmund Gartner, Naval Postgraduate School