Carthago Delenda Est: Co-opetitive Indirect Information Diffusion Model for Influence Operations on Online Social Media

TL;DR

This work introduces Diluvsion, an agent-based model for simulating bot-driven information operations on Twitter-like platforms, emphasizing co-opetitive diffusion, indirect influence, and theme-level dynamics. By integrating stances, themes, memory, engagement signals, and non-social tie transmission within a 30-minute resolution, the model reproduces realistic diffusion patterns and enables testing of both orthodox and adversarial info ops strategies. The authors validate the approach against real-world data and demonstrate key findings on polarization, bot impact, and theme propagation, including a notable result where theme conservation can spread without shifting overall stance distributions. The framework offers a practical tool for planning, defense, and policy analysis in the context of decentralized, led-by-bots information campaigns and highlights the nuanced role of engagement cues and non-tie information pathways in shaping public discourse.

Abstract

For a state or non-state actor whose credibility is bankrupt, relying on bots to conduct non-attributable, non-accountable, and seemingly-grassroots-but-decentralized-in-actuality influence/information operations (info ops) on social media can help circumvent the issue of trust deficit while advancing its interests. Planning and/or defending against decentralized info ops can be aided by computational simulations in lieu of ethically-fraught live experiments on social media. In this study, we introduce Diluvsion, an agent-based model for contested information propagation efforts on Twitter-like social media. The model emphasizes a user's belief in an opinion (stance) being impacted by the perception of potentially illusory popular support from constant incoming floods of indirect information, floods that can be cooperatively engineered in an uncoordinated manner by bots as they compete to spread their stances. Our model, which has been validated against real-world data, is an advancement over previous models because we account for engagement metrics in influencing stance adoption, non-social tie spreading of information, neutrality as a stance that can be spread, and themes that are analogous to media's framing effect and are symbiotic with respect to stance propagation. The strengths of the Diluvsion model are demonstrated in simulations of orthodox info ops, e.g., maximizing adoption of one stance; creating echo chambers; inducing polarization; and unorthodox info ops, e.g., simultaneous support of multiple stances as a Trojan horse tactic for the dissemination of a theme.

Figures (9)

• Figure 1: At time-step 1, a neutral human agent (H1) makes a tweet (action not depicted in this figure) expressing a neutral stance and themes of medical authority and individual liberty over collective good, e.g., "My doctor tells me wearing masks is a personal preference." (A) While still at time-step 1, a second human agent (H2), H1's follower, concocts a reply-tweet with matching themes but a positive stance, e.g., "The CDC said masks protect everyone, not just the wearer." (B) A negative bot (B1) calls poll_tweets, which is equivalent to doing a search of existing tweets, finds H1's neutral tweet, and replies with a negative tweet. (C) A third human agent (H3) makes a reply-tweet with a negative stance to H2's earlier reply to H1. Two bots (B2 and B3) liked H3's reply-tweet. H3's reply happens to be seen by H1 because H1, as a follower of H2, occasionally reads the replies to H2's tweets. (D) A fourth human agent (H4) makes a tweet with a positive stance that is seen by H1, which pushes the very first tweet seen out of H1's limited memory. (E, not shown) At the start of time-step 2, when read_conversion is called, H1 observes one positive-stanced tweet without any engagement, two negative-stanced tweets, one which received two likes, and finally its own neutral-stanced tweet that received three replies. To determine if H1 will experience a stance conversion after reading these tweets, we check if a generated uniform random number clears H1's resistance threshold. The number does. To see which stance from the seen tweets infect H1, a weight is calculated for each stance. The negative stance has a heavier weight than neutral or positive because the negative tweets are greater in number and have affirmative engagement instead of negatory (likes instead of replies). Weighted random choice picks the negative stance, H1 changes its stance to negative.
• Figure 2: Engagement metrics: real-world data vs simulations with 1,000 agents.
• Figure 3: Temporal evolution of stance distribution among 1,000 agents. Values are from 20-run averages of different scenarios/simulations. Values at simulation end corresponds to the mean values reported in Table \ref{['tab:simulations']}. Descriptions of the scenarios are found in Table \ref{['tab:simulations']} and in Section \ref{['sec:simulations']}.
• Figure 4: Temporal evolution of mean follower and followee (source) counts split by an agent's stance for 20-run averages of different scenarios/simulations. When an agent converts to a stance at a time-step, its follower and source counts are included in the average for the new stance and excluded for the old stance. Subplots do not share a common range for the y-axis due to significant variance across scenarios. Descriptions of the scenarios are found in Table \ref{['tab:simulations']} and in Section \ref{['sec:simulations']}.
• Figure 5: Embedded bots scenario. Each edge arrow represents the direction of information flow, i.e., the inverse of a directed follower-followee relationship, for a simulation with 100 agents in conditions equivalent to Sim 13 (Table \ref{['tab:simulations']}). In standard graph notation for an edge/link $(u, v)$, $u$ is a follower of $v$, so the follower-followee relationship is $u \rightarrow v$ but the information flow is reversed, $u \leftarrow v$. The image on the left shows the network at simulation start and has nodes sized according to the ratio of followers to followees. The image on the right shows the graph at simulation end and nodes are sized according to the average ratio of likes to replies received by the agent's tweets $\frac{1}{N} \sum_{N}^{i=1}\frac{\# \text{likes}_i}{\# \text{replies}_i}$, with higher values indicating that the agent's tweets tend to have a perception of being well-received. As likes and replies are accumulated over the course of a simulation, sizing the nodes based on engagement metrics is only possible post-simulation. The edge/link widths for the image on the left has a fixed value, while those on the right are based on stance and themes compatibility between two agents. Stance is indicated by each node's color. Agency is indicated in the edge/link color and the node's border color. Blue nodes are the opposition, yellow nodes are neutral, and green nodes are supporters. A red border indicates a bot while a white/invisible border indicates a human.