Frequency-Domain Characterization of Memory Poisoning Propagation in Multi-Agent Collaborative Systems

Abstract

Memory poisoning in multi-agent collaborative systems may exhibit structured propagation patterns across communication channels. This study analyzes poisoning diffusion using frequency-domain decomposition of inter-agent memory update signals. Memory state transitions are transformed via discrete Fourier analysis to identify anomalous high-frequency components introduced by adversarial perturbations. A spectral attenuation filter selectively suppresses abnormal frequency bands before synchronization. Experiments were conducted on 120-agent distributed coordination tasks over 10,000 update cycles. Poisoning injection at 15% intensity increased high-frequency spectral energy by 63.2% compared to clean baselines. Spectral filtering reduced contamination spread from 58.4% to 19.6% of agents and restored task performance to 87.3% of nominal levels. Frequency-domain modeling provides a novel perspective on poisoning detection and containment in collaborative agent systems.