Introduction
Distributed systems, ranging from smart grids to swarms of autonomous drones, rely on "average consensus"—the ability of all agents to agree on a common value based on their collective local data. However, these systems are notoriously fragile. A single "misbehaving" agent, whether due to a cyberattack or a hardware fault, can inject biased or noisy data that "drags" the entire network away from the truth. This paper explores whether a controlled amount of competition can actually enhance the resilience of these cooperative networks.
The Resilience Gap in Sparse Networks
Existing solutions usually rely on filtering strategies like Weighted Mean Subsequence Reduced (W-MSR), where agents ignore the most extreme values they receive from neighbors. While effective, these methods have a major limitation: they only work if the network is highly connected (specifically, "r-robust"). In many real-world, sparse networks, these guarantees vanish, leaving the system vulnerable. The authors propose a simpler, more robust way to achieve resilience without complex filtering or global knowledge of network topology.
Methods: The Friedkin-Johnsen Model
The authors draw inspiration from social science, specifically the Friedkin-Johnsen (FJ) model of opinion dynamics. In this model, agents don't just blindly average their neighbors' opinions; they maintain a degree of "prejudice" or "stubbornness" toward their initial observation.
The key innovation is the introduction of a competition parameter ($\lambda$):
- $\lambda = 0$ (Full Collaboration): Standard consensus. Agents trust their neighbors completely.
- $\lambda = 1$ (Full Competition): Agents ignore everyone else and stick only to their own data.
- $\lambda \in (0, 1)$: A hybrid approach where agents "anchor" to their own data while still listening to the network.
The Competition-Collaboration Trade-off
The paper's central insight is that resilience isn't a binary state. Instead, there is a fundamental trade-off. Pure collaboration leads to "deception" by bad actors, while pure competition leads to "inefficiency" because agents don't benefit from the network's collective knowledge.
As illustrated in the figure above, the total network cost ($e_{tot}$) is the sum of deception error ($e_{deception}$) and consensus error ($e_{consensus}$). While deception error drops as $\lambda$ increases (more stubbornness), the consensus error rises. The optimal strategy is almost always a hybrid, represented by the minimum of the $e_{tot}$ curve.
Key Results and Performance
The authors mathematically characterize the consensus error and prove that there is a non-trivial optimal value for $\lambda$ ($\lambda^*$) that minimizes the damage caused by misbehaving agents.
Optimal Stubbornness and Attack Intensity
Numerical tests show this "sweet spot" is often quite small (around $\lambda \approx 0.1$ to $0.2$), meaning even a little bit of stubbornness goes a long way. Furthermore, the optimal level of competition is sensitive to the threat level.
As the number of malicious agents ($M$) or the intensity of noise ($d$) increases, the optimal $\lambda^*$ also increases. This suggests that agents can adapt their level of "stubbornness" based on how much they suspect the network is being compromised.
Neutralizing Social Power
The authors use the concept of "social power" to explain how influence spreads. They show that anchoring to local data effectively "dilutes" the social power of a malicious agent. In random geometric graphs, "hubs" (highly connected nodes) can be a liability if compromised. The FJ model's anchoring mechanism is particularly effective at neutralizing this "hub-based" spread, preventing an attacker from dominating the network's final state.
Limitations and Future Work
While promising, the approach has limitations. Finding the exact optimal $\lambda$ currently requires some knowledge of the adversary's noise and bias levels, which might not be known in advance. Future work could explore heterogeneous competition, where different agents have different levels of stubbornness based on their own local confidence or trust scores. Additionally, "smart" adversaries who actively try to hide their presence would require a more complex game-theoretic analysis.
Summary updated