Multi-Agent Adversarial Synthesis and Defense Pipeline
Figure 2: Overview of the proposed framework for multi-agent adversarial synthesis and defense benchmarking. The pipeline is divided into (Left) a training data generation phase using varied network topologies (Chain, Tree, Random) to produce text and numerical embeddings, and (Right) a defense evaluation phase where models are benchmarked through iterative rounds of live debate and malicious agent pruning.
Paper context
Paper title: GAMMAF: A Common Framework for Graph-Based Anomaly Monitoring Benchmarking in LLM Multi-Agent Systems Abstract: The rapid integration of Large Language Models (LLMs) into Multi-Agent Systems (MAS) has significantly enhanced their collaborative problem-solving capabilities, but it has also expanded their attack surfaces, exposing them to vulnerabilities such as prompt infection and compromised inter-agent communication. While emerging graph-based anomaly detection methods show promise in protecting these networks, the field currently lacks a standardized, reproducible environment to train these models and evaluate their efficacy. To address this gap, we introduce Gammaf (Graph-based Anomaly Monitoring for LLM Multi-Agent systems Framework), an open-source benchmarking platform. Gammaf is not a novel defense mechanism itself, but rather a comprehensive evaluation architecture designed to generate synthetic multi-agent interaction datasets and benchmark the performance of existing and future defense models. The proposed framework operates through two interdependent pipelines: a Training Data Generation stage, which simulates debates across varied network topologies to capture interactions as robust attributed graphs, and a Defense System Benchmarking stage, which actively evaluates defense models by dynamically isolating flagged adversarial nodes during live inference rounds. Through rigorous evaluation using established defense baselines (XG-Guard and BlindGuard) across multiple knowledge tasks (such as MMLU-Pro and GSM8K), we demonstrate Gammaf's high utility, topological scalability, a Passages referencing this figure: of adversarial agents. K eywords Large Language Model ⋅ \cdot Multi-Agent System ⋅ \cdot Anomaly Detection ⋅ \cdot Graph Neural Network ⋅ \cdot Adversarial Agents ⋅ \cdot Benchmarking Framework ⋅ \cdot Defense Mechanism ⋅ \cdot Network Topology ⋅ \cdot Attack Detection ⋅ \cdot Agent Collaboration ⋅ \cdot Prompt Injection ⋅ \cdot Security ⋅ \cdot Topological Defense ⋅ \cdot LLM-MAS 1 Introduction Figure 1: Example of debate setup for collaboration in a LLM-MAS. Agents exchange natural language discourse to reach a consensus on a specific task. The diagram illustrates how the communication structure constrains information flow, requiring agents to synthesize the logical reasoning of their neighbors to update their internal context. The increase in performance of large language models (LLMs) ted with external tools and retrievable memory, enhancing their reasoning abilities and enabling them to observe and act on their environment. These LLM-powered agents have demonstrated success over isolated language models and enable more dynamic functionalities (Mialon et al. , 2023 ) . The next step in development has been the integration of multiple agents to collaborate together, as shown in Figure 1 , where agents debate with each other to reach a common solution. Studies show that combining agents in a properly configured communication network increases the likelihood of successful task completion, especially for more complex tasks ( Talebirad and Nadiri ( 2023 ) ; Qian et al. ( 2023 ) ). Moreover, the adoption of multi-agent systems (MAS) is not limited to academic research; they a
