a brief conflict modeling primer

In an attempt to keep myself accountable regarding my thesis and spring-term research, here are some interesting updates on what the US government is doing regarding conflict forecasting and risk assessment.

Over the last two and a half (whew, already?) months at the Office of Science and Technology Policy, I have learned about, among many other things, the behemoth undertaking that is the federal AI use case inventory.

I’ll save you the time of scraping through the 700 lines of use cases released by federal agencies in 2023, though they can be found here if you are curious. Upon first glance (and then many, many more run-throughs), there were three projects released by the Department of State that particularly piqued my interest regarding peacekeeping and conflict monitoring.

Mainly supported by the Bureau of Conflict and Stabilization Operations (CSO), these projects are titled:

1. Verified Imagery Pilot Project

2. Conflict Forecasting

3. Automated Damage Assessments (in conjunction with the Conflict Observatory)

What follows is a recounting of each use case and my personal opinions on the viability and existing work surrounding each topic.

[Aside: in the middle of writing this piece, on November 8th, CSO announced a new initiative in conjunction with 8 universities called the Academic Centers of Conflict Anticipation and Prevention (ACCAP), further evidence that the attention and funding being turned onto conflict modeling is, hopefully, growing.]

The official State description is as follows:

The Bureau of Conflict and Stabilization Operations ran a pilot project to test how the use of a technology service, Sealr, could verify the delivery of foreign assistance to conflict-affected areas where neither U.S. Department of State nor our implementing partner could go. Sealr uses blockchain encryption to secure photographs taken on smartphones from digital tampering. It also uses artificial intelligence to detect spoofs, like taking a picture of a picture of something. Sealr also has some image recognition capabilities. The pilot demonstrated technology like Sealr can be used as a way to strengthen remote monitoring of foreign assistance to dangerous or otherwise inaccessible areas.

In recent years, there’s been quite a bit of noise around using blockchain encryption to monitor digital tampering. Most notably, I am reminded of Stanford and USC’s Starling Lab, a research lab that attempts to use open-source frameworks and decentralized authentication techniques to verify digital media. Hala Systems, which I wrote my last post on, has also been known to use blockchain methods in this space. The two organizations have been working together to put together a cryptographic dossier of possible war crimes in Ukraine, for example.

In case you want to dive deeper: Just two weeks ago, CSIS published a comprehensive walkthrough on the use of blockchain in strengthening information integrity in democracies. It’s quite comprehensive and also details how Reuters has been partnering with Canon to equip professional photojournalists with a camera that registers images immediately to the public ledger upon creation. (Almost a bit like the recent work released by Deepmind on watermarking.)

I’m generally quite interested in this information management tactic, though it doesn’t necessarily address the whole problem. For example, how does one manage what happens to a digital asset before it is loaded onto a shared ledger? Regardless, information integrity isn’t really conflict forecasting and I could write a whole other post on this alone. Still, I do count it broadly as part of conflict mitigation (see: Myanmar circa the early 2010s) and am curious to know the results of the CSO pilot project—as of now, there is no additional news from CSO regarding the viability of Sealr in real conflict settings.

The official State description:

CSO/AA is developing a suite of conflict and instability forecasting models that use open-source political, social, and economic datasets to predict conflict outcomes including interstate war, mass mobilization, and mass killings. The use of AI is confined to statistical models including machine learning techniques including tree-based methods, neural networks, and clustering approaches.

I’m planning on doing a more comprehensive scan of the conflict forecasting research landscape (see my work in progress here), but from anecdotal observations, the State project seems to closely reflect the existing work on modeling. Minus the use of GIS data, the project covers three of the most common approaches to modeling conflict.

My best guess at the datasets they might be using includes the big three:

• ACLED: Armed Conflict Location & Event Data Project

• UCDP: Uppsala Conflict Data Program

• ViEWS: Violence & Impacts Early-Warning System by PRIO and Uppsala

• and some aggregation of localized social/news media

To date, I’ve seen very few other attempts at creating a suite of open-source modeling—particularly ones that serve to aggregate different models for the use of a non-technical third party. Many have worked on ensemble models and/or layering human expertise to better capture critical domain expertise (for example: AutoML, published by D’Orazio et al. in 2019), but the only actual open-source tool that comes to mind is CoPro, published by Jannis Hoch et al. in 2021. As such, I really do applaud this effort and am excited to see where CSO takes it.

• Fun fact, Hoch’s second paper is what I built v0 of my conflicts research network analysis (link above) off of.

More fringe: In other modeling news, a recent WIRED article has gotten me thinking about simulations. It’s one of the more commercializable and non-traditional methods of conflict modeling and, to be honest, I don’t know enough about it to take it with more than a grain of salt. Still an interesting read and something to look out for:

• WIRED released an article on 11.02.2023 discussing the modeling of conflicts (by a company called CulturePulse) via digital, AI-generated population copies.

  • Which then, in turn, reminds me of what Devant is doing in synthetic data generation via digital humans.

• Insofar as I can tell, public government efforts on this front focus more on the logistics of waging warfare (see: JCATS). Similar approach, but pretty orthogonal in objectives.

This is the sub-field of conflict research I’m most excited about—there is a sneakily large body of work out there on algorithmic approaches to forecasting. At the same time, I think that the criticism leveled towards conflict modeling is generally warranted—so much contextual, boots-on-the-ground knowledge about lived experiences cannot be captured (yet or ever?) in this ten-thousand-foot-view kind of work. Nevertheless, I remain steadfast in my belief that initiatives like this (and, hopefully, ACCAP) are not perfect, but also certainly not wasted.

State description:

The Conflict Observatory program uses AI and machine learning on moderate and high-resolution commerical satellite imagery to document a variety of war crimes and other abuses in Ukraine, including automated damage assessments of a variety of buildings, including critical infrastructure, hospitals, schools, crop storage facilities.

Since this is housed mainly at the Conflict Observatory and is less directly under the purview of State, I’ll skip over this one for now. The work at Conflict Observatory could (and should) be its own separate post—I don’t think I can do it justice here—but I wanted to mention it as a CSO-reported use case.

Broadly, looking at the Department of State’s AI use case inventory actually gives a great overview of the three prongs of conflict-related research that’s out there, as of now: 1) blockchain in information integrity, 2) machine learning in conflict modeling, and 3) OSINT in conflict monitoring.

As a European told me earlier this fall, a lot of what the Americans do on this front might very well be obscured by our large security apparatus, which precludes the general public from learning more about this field. However, as people in and around conflicts continue to generate new monitoring and data-collection tools, the civilian conflict research field grows more promising by the year. Still, I do wonder what parts of government are primarily investing in this—I’m not sure to what degree the answer is the Department of State [edit: this sentence was written before the announcement of ACCAP], but CSO is doing a great job of grabbing a small piece of the limelight for the field and I’m incredibly excited to see how their work unfolds in the coming months.