FedAHPIP: Federated Learning with Adaptive Hot Parameter Identification and Personalized Anchoring for multi-agent collaboration

Smart manufacturing seeks to achieve collective intelligence through collaboration. However, such collaboration must be secure and personalized to handle heterogeneous industrial agents. Federated learning offers a promising paradigm for this setting but faces two fundamental challenges: privacy leakage through gradient inversion attacks (e.g., DLG) and data heterogeneity requiring personalized models. To address these challenges, we propose FedAHPIP, a federated learning framework that integrates secure aggregation with personalized learning. Our approach includes an adaptive hot parameter identification mechanism that dynamically identifies sensitive parameters (hot parameters) based on their update momentum, layer semantics, and potential label leakage risks. By focusing encryption on these hot parameters, FedAHPIP drastically reduces the privacy leakage surface. We also develop a personalized anchoring strategy that allows each agent to retain its critical parameters while assimilating knowledge from the global model, effectively balancing personalization and collaboration. Extensive experiments on benchmark and industrial datasets demonstrate that FedAHPIP achieves superior personalized accuracy under extreme non-IID settings, provides robust security against DLG attacks, and maintains minimal computational overhead. FedAHPIP thus offers a practical solution for trustworthy collective intelligence in smart manufacturing environments.