Reconnaissance of Influence Operations

Researchers at MIT Lincoln Laboratory demonstrated an end-to-end software prototype that automates the detection of hostile influence narratives, the identification of accounts engaged in influence operations, and the quantification of the impact individual actors have on spreading their disinformation narrative. The system combines advanced artificial intelligence techniques: natural language processing, machine learning, graph analytics, and a causal inference approach to measuring influence within a network.

Motivation

Using disinformation and propaganda as tools of warfare dates back to antiquity. However, the contemporary information environment powered by the Internet enables hostile actors to conduct such campaigns with immense scale, speed, and reach. These influence operations have been employed against the United States and its allies in attacks on their democratic processes. Recently, widespread influence operations undertaken by hostile foreign actors have spread misinformation about the COVID-19 pandemic, threatening public health.

Methodology

1. Collect potentially relevant social media content (e.g., Twitter posts) based on keywords, accounts, languages, and spatiotemporal ranges.
2. Identify distinct narratives by using a type of machine-learning-based natural language processing called topic modeling.
3. Tag the accounts related to narratives identified in step 2 with a machine-learning classifier score based on their behavioral, linguistic, and content features. Steps 2 & 3 may be iterated to provide a more focused corpus for IO narrative detection.
4. Use graph analytics to construct an influence operations (IO) social network based on the pattern of interactions of accounts.
5. Quantify the impact of each account — measured by its contribution to the narrative spread over the network — by using a novel network causal inference methodology.

The end product of this methodology is a mapping of the IO narrative network in which IO accounts of high impact are identified. This approach has measurable and noteworthy advantages over traditional impact statistics that are based on activity count (e.g., tweet and retweet counts) and network topology (e.g., network centralities) in discovering high-impact IO accounts.

Additional Resources

Patent pending, 2019/0311441 A1

More Information

S.T. Smith et al., "Automatic Detection of Influential Actors in Disinformation Networks," Proceedings of the National Academy of Sciences of the United States of America, 26 Jan. 2021.

2020 R&D 100 Award winner