Fake News in Social Networks

TL;DR

The paper develops a deep multi-agent reinforcement learning framework to model fake-news diffusion as an information-aggregation game on social networks, incorporating an adversary that biases private signals. Through independent Q-learning in both simulated and human-subject experiments, it shows that attackers targeting high-degree nodes and users with weak private information substantially impair aggregation, with spread-out biases across clusters being more effective than concentrated attacks. Citizens can learn to adapt to attacks, albeit at a cost to baseline accuracy, and network topology (rewired vs clustered) modulates vulnerability. The work provides actionable insights for countermeasures, such as balancing networks and preserving privacy of connectivity, and introduces a flexible methodology for examining misinformation dynamics in complex networks.

Abstract

We propose multi-agent reinforcement learning as a new method for modeling fake news in social networks. This method allows us to model human behavior in social networks both in unaccustomed populations and in populations that have adapted to the presence of fake news. In particular the latter is challenging for existing methods. We find that a fake-news attack is more effective if it targets highly connected people and people with weaker private information. Attacks are more effective when the disinformation is spread across several agents than when the disinformation is concentrated with more intensity on fewer agents. Furthermore, fake news spread less well in balanced networks than in clustered networks. We test a part of our findings in a human-subject experiment. The experimental evidence provides support for the predictions from the model, suggesting that the model is suitable to analyze the spread of fake news in social networks.

Figures (8)

• Figure 1: Left: Instance of a Barabsi-Albert random network with $|N| = 10$ that was used in our analyses (unless otherwise stated). Middle: stereotypical clustered network with $|N|=12$ with three ($n_c=3$) fully connected clusters of four nodes ($s_c=4$) each. Right: the clustered network in balanced form produced by random rewiring of link pairs (preserves degree distribution). These networks were used in our analysis of the effectiveness of information aggregation and attack (i) in clustered vs. balanced networks and (ii) for spread out vs. focused attack strategies.
• Figure 2: (A-B): Information aggregation over time in a Barabasi-Albert graph. (A) Baseline scenario. Under the attack test scenario, a single agent receives a biased signal. (B) Biased signal attack scenario. The upper and lower dashed lines corresponds to the benchmarks $\mathbb{E}[\mathds{1}\{\hat{\theta} = \theta\}]$ and $\mathbb{E}[\mathds{1}\{\tilde{\theta} = \theta\}]$ respectively. Square markers correspond to the accuracy achieved by the De Groot information aggration heuristic in the baseline and biased signal attack scenarios respectively.
• Figure 3: (A-B) Agents are trained in the absence of any attacker on a Barabasi-Albert graph. (A) Top: Average decline in accuracy conditional on agent id which determines network position. Bottom: degree of attacked agent. (B) Effectiveness of conditioning on attacked agent signal strength. (C) Agents trained in the absence of any attacker for clustered (left) and balanced (right) graphs. Base: no attack. Focus: a single, randomly chosen agent receives a strong biased signal ($\beta^i=3$). Spread: two, randomly chosen agents receive weak biased signals ($\beta^i=1.5$). Blue markers represent training runs with different seeds for a given network and attack scenario. Orange markers are averages over the different training runs. Error bars computed over outcomes with different random seeds but fixed neural network weights are too small to be visible on the graphs.
• Figure 4: Rewired network (left) and clustered network (right), as used in the experiment.
• Figure 5: Mean accuracy in the treatments of the experiment.

Theorems & Definitions (2)

• Definition 1: Information aggregation
• Definition 2: Accuracy of information aggregation