Abstract
Collaborative agents frequently rely on shared memory states without external supervision. This study proposes a self-supervised consistency regularization mechanism to detect and correct poisoned memory updates. Each agent maintains a local predictive model to estimate expected peer memory updates. Deviations are penalized using a consistency loss computed over sliding windows of interactions. The framework was evaluated on a 75-agent distributed planning benchmark under poisoning rates ranging from 5% to 25%. At a 20% injection rate, baseline systems experienced a 36.7% performance drop, whereas consistency-regularized systems limited degradation to 14.2%. False-positive rejection rate remained below 7.1%. Self-supervised consistency constraints enhance collaborative robustness without centralized validation.

This work is licensed under a Creative Commons Attribution 4.0 International License.
Copyright (c) 2026 Jeroen van Dijk, Maarten de Vries, Sophie van der Meer (Author)