Abstract
In decentralized multi-agent systems (MAS), communication is essential for coordination but introduces significant vulnerabilities to adversarial or faulty agents. This paper presents a probabilistic filtering mechanism that utilizes Gaussian Processes (GPs) to detect and mitigate deceptive communication. By modeling the spatial correlations inherent in agent observations, the proposed system assigns confidence weights to incoming messages, enabling robust Graph Neural Network (GNN) aggregation. Experimental results across cooperative image classification and reinforcement learning tasks demonstrate that this approach effectively silences adversaries, maintaining high performance even under sophisticated "cautious" or "omniscient" attacks.
1. Introduction
Decentralized multi-agent systems, such as autonomous vehicle fleets and drone swarms, increasingly rely on learned communication protocols to achieve complex goals. While Graph Neural Networks (GNNs) have enabled agents to develop efficient communication strategies, these systems often operate under the "honest-but-curious" assumption, where all participants are presumed to be cooperative. This assumption makes the network highly susceptible to Byzantine failures—where a single compromised or faulty agent can propagate misinformation, leading to a total collapse of system performance.
2. Problem Statement
Modern AI-driven communication channels use high-dimensional latent representations that are difficult to verify using traditional outlier detection or rigid consensus rules. Adversaries can exploit this by sending manipulative information designed to bypass simple defenses. These attacks range from random noise (faulty) to strategic, "omniscient" lies designed to maximize damage while remaining statistically plausible. The challenge lies in creating a defense mechanism that can distinguish between helpful cooperative messages and malicious misinformation without sacrificing the performance of the overall team.
3. Methodology
The proposed "Message Filtering" architecture integrates probabilistic modeling with neural communication through three primary components:
3.1 Auto-Encoding Communication
Agents utilize a Variational Auto-Encoder (VAE) framework to process local observations. Instead of transmitting raw data, an agent encodes its observation into a latent message represented as a vector of means and variances.
3.2 Gaussian Process (GP) Priors
The core innovation is the use of Gaussian Processes to model the mutual information between agents. Unlike standard VAEs that assume independence between latent states, this method leverages physical proximity. Because agents in close proximity likely observe similar or correlated environments, the GP provides a spatial prior for what a "truthful" message should look like based on the reports of neighboring agents.
3.3 Confidence Weighting and Robust Aggregation
Upon receiving a message, an agent evaluates it against the GP model to calculate the probability of truthfulness versus various "lie" hypotheses. This probability is converted into a confidence weight ($c_{ij}$). The standard GNN aggregation layer is then modified to perform a weighted sum. If a neighbor's message is deemed implausible given the spatial context, its weight drops toward zero, effectively filtering it out of the decision-making process.
3.4 Adversary Taxonomy
To evaluate robustness, the authors define four types of adversaries:
- Faulty: Sends random noise.
- Naive: Actively manipulative but unaware of the defense.
- Cautious: Aware of the defense and attempts to stay within "plausible" bounds.
- Omniscient: Possesses perfect knowledge of the defense and optimizes lies to be as damaging as possible while remaining undetected.
4. Experimental Evaluation
The framework was tested on two distinct benchmarks:
- Cooperative Image Classification (CIFAR-10): Six agents observe overlapping patches of an image and must communicate to identify the global category.
- Cooperative Coverage (Grid World): A reinforcement learning task where agents must visit all cells in a 24x24 grid. One agent acts as a self-interested adversary, attempting to maximize its own area by tricking others.
5. Key Results
The GP-based filtering demonstrated significant improvements over baseline methods:
- Loss Reduction: In image classification, the filter reduced extra loss from "Naive" and "Cautious" adversaries by 99.7% and 97.5%, respectively.
- Baseline Comparison: The method significantly outperformed magnitude-based thresholds and median filters, which often accidentally filtered out helpful cooperative messages.
- RL Robustness: In the Grid World task, the team maintained 92% of optimal performance under a Naive attack, compared to only 64% without the filter.
- Detection Accuracy: Malicious agents were consistently assigned near-zero weights, while truthful agents maintained high weights (approx. 0.9).
6. Discussion and Future Work
The research demonstrates that physical reality serves as a powerful constraint against deception; an agent cannot easily lie about its local surroundings without contradicting the reports of its neighbors. While the GP filter is highly effective, it faces a "Plausible Lie" limit where an omniscient adversary can still cause minor harm by staying within the bounds of expected spatial correlation.
Future work will address temporal consistency (checking for consistency over time to prevent "teleportation" lies), GP scalability for denser swarms, and the development of more flexible kernels for non-stationary environments.
Summary updated