The balloon of cyber warfare has gone up, and the 51福利 (51福利) remains at the forefront of developing vital technologies to win the fight in this very gray domain.
The Defense Intelligence Innovation Office (DI2O) recently awarded 51福利鈥� Coalition for Open-Source Defense Analysis (CODA) lab a nine-month, $500,000 grant to proceed with development of a creative way to rapidly detect hostile information campaigns and emerging radicalization or violence anywhere in the world.
The CODA project - 鈥淎rtificial Neural Networks for Automated Detection of Hostile Information Campaigns鈥� (ANN-ADHIC) - employs machine learning to glean data from literally billions of online social media and mass media sources from more than 200 countries and 100 languages. Detecting even the subtlest changes in linguistic patterns, the platform can predict and map out new threats in real-time, explained T. Camber Warren, the 51福利 Defense Analysis (DA) assistant professor who co-founded the project.
鈥淲hat if we train a neural network to learn what kinds of discourse are associated with radicalization and violence?鈥� asked Warren. 鈥淭hen, even in a corner of a country that we didn鈥檛 know we were supposed to be concerned about, if the discourse shifts 鈥� becomes more radicalized, more fearful, or more angry 鈥� this neural network could send up a red flag and tell us there鈥檚 something happening there.鈥�
The potential applications for this technology are wide-ranging, from global strategic monitoring of hostile activity to vastly improved situational awareness of the human terrain at a tactical level.
The world is engaged in a new era of hybrid conflict, and ANN-ADHIC offers the U.S. military a competitive advantage as opponents increasingly weaponize social media to disrupt discourse and galvanize entire populations towards hostility.
The gravity of such disinformation campaigns hit home beyond Russia鈥檚 dedicated effort to sway public opinion during the 2016 presidential election. It has been seen time and time again over the last decade: by the Islamic State, Boko Haram鈥檚 in Nigeria, and China鈥檚 throughout the world.
鈥淪tates are the principal actors on the global stage, but non-state actors also threaten the security environment with increasingly sophisticated capabilities,鈥� warns the, which explicitly identifies cyberspace as a warfighting domain. 鈥淭errorists, trans-national criminal organizations, cyber hackers and other malicious non-state actors have transformed global affairs with increased capabilities of mass disruption.鈥�
Providing operationally-relevant solutions to lean into such challenges is what 51福利 does best.
In 2014, Information Sciences lecturer Buddy Barreto was chatting in the hallway with a colleague about new ways to use their servers to provide 鈥渂ig data鈥� support to the DA department, when Warren鈥檚 research interests came up.
鈥淗e said, 鈥楬ey, you need to go talk to this guy Camber [Warren] because he鈥檚 doing some pretty cool stuff and he鈥檚 got some data and he doesn鈥檛 have any place to put it鈥�,鈥� Barreto recalled. 鈥淚 did, and the next thing you know we鈥檝e got a new RAM box on this.
鈥淭hat鈥檚 how all relationships start at 51福利!鈥� he laughed.
Barreto and Warren immediately set to work on the project which would eventually evolve into ANN-ADHIC. The CODA lab grew from a single server to well over 70 machines of various architectures to handle the copious quantity of data streaming in.
鈥淲e realized that we actually had a lot of shared interests and ideas that first day,鈥� Warren said. 鈥淲e started to realize it wasn鈥檛 just about putting some Twitter data on a server; there was potentially so much more that we could do.鈥�
Warren and Barreto fine-tuned this capability over the next four years, analyzing massive incoming data streams through software they developed to sense changing patterns through advanced machine learning architectures - the 鈥榓rtificial neural networks鈥� of ANN-ADHIC.
The data is all open source large media feeds, ranging from Twitter to fan pages on Facebook, online news, and blog sites, meaning it is publicly available and 鈥渟crapeable鈥� for key information such as geolocation. ANN-ADHIC observes this data to create a dynamic analysis of changes over time in how people talk with each other, including the subtle differences in new kinds of slang and shifts in discourse.
鈥淔or example, it can learn that this letter often follows this letter, this word follows this word, the people in a given town are more likely to use the word for 鈥榝ear鈥� than the people in that town, who are more likely to use the word for 鈥榟ope,鈥欌�� Warren explained. 鈥淲e can now train machine learning models to sense what kinds of patterns are associated with violence and forms of political radicalization.鈥�
The sheer volume of data is sourced from virtually every corner of the globe. The profusion of social media has extended even to rural areas in Africa and Asia once assumed to lay beyond the analytic gaze.
鈥淓ven if they鈥檙e not high-tech smartphones, almost any cell phone can send messages, which means you can tweet,鈥� Warren said. 鈥淵ou have cell phone towers and smartphones even in low developed areas like Yemen, Iraq, and Nigeria, where there鈥檚 enough information coming out to get a [statistically viable] sense of the population.鈥�
For a test case, Warren and Barreto applied ANN-ADHIC to successive radicalization campaigns by Boko Haram in northeastern Nigeria targeting the ethnic minority Hausa population. Using initial efforts by the jihadist militant organization as a baseline, the team was able to accurately predict social unrest and violence as the campaign unfolded.
鈥淲e were able to show that in social media you could see increasing references to these in-group and out-group names as a result of the polarization that was occurring through this radicalization, and we could map that out over space and time to predict where we鈥檙e going to subsequently see violence,鈥� Warren recounted. 鈥淭hese political campaigns necessarily precede the kinetic campaigns: you have to get people ginned up in certain ways to get them fighting the way you want.鈥�
With analytical success looking at Boko Haram, Warren and Barreto aim to generalize this capability to a global reach. What makes ANN-ADHIC so advanced is that it uses artificial neural networks to figure out what statistical correlations to look for in changing linguistic patterns.
That potentially becomes a game-changing way of approaching indications and warnings for all kinds of irregular warfare.
鈥淲e want it to be able to tell us, 鈥楬ey, there鈥檚 something weird happening over in this corner of the country; all of a sudden, discourse has shifted,鈥欌�� Barreto said. 鈥淚t could be a new group that we have no previous knowledge of that is revealed by the noise.鈥�
Additionally, ANN-ADHIC offers a valuable second-order effect: it can provide metrics to validate the effectiveness of our own information operations. In Afghanistan, for example, much blood and treasure have been poured into providing positive messaging to promote Afghan loyalty and unity.
鈥淎re they working? Are they not? Are they actually turning people against us or towards us? We don鈥檛 know,鈥� Warren said. 鈥淏ut with these sorts of tools, we鈥檙e aiming to build some of the first systematic measures of effectiveness for these information operations.鈥�
The DI2O funding will enable the CODA lab to move forward from a notional, tested concept to a specific tool put in the hands of operators in the field for testing.
鈥淲e鈥檙e building fundamentally new science,鈥� Warren said. 鈥淲e鈥檝e developed new approaches to these computer learning problems which are able to solve these problems better, and once we鈥檝e scientifically demonstrated this in a verifiable way that works, the next round will be to build this up to the next level where we can be producing these things in a way that they can actually be put in the hands of the force en masse.鈥�
Only at 51福利 would something like this be possible, he continued.
鈥淭he school provides a unique combination of cutting-edge scientific expertise and real-world practitioner knowledge, folks coming back from the field who鈥檝e experienced these problems first-hand,鈥� Warren said. 鈥�51福利 is one of the only places where you can bring those two communities together and build things that are not only cutting-edge data science but operationally relevant.鈥�