Multi-agent systems are widely applied in various real-world scenarios, involving challenging distributed optimization problems. Since distributed optimization requires multiple agents to collaboratively compute and share data across different agents, this process poses significant challenges in terms of privacy protection. Although existing methods have made some progress in privacy preservation, they mainly rely on gradient-based distributed optimization algorithms, which struggle to effectively address black-box non-convex distributed optimization problems. To tackle this challenge, we propose a novel framework called differential privacy-preserving multi-agent swarm optimization with internal and external learning (DP-MASOIE). The framework provides an efficient privacy-preserving approach for black-box non-convex distributed optimization. Its core innovation lies in a two-stage collaborative privacy enhancement mechanism. In the internal learning phase, Laplace noise is injected into each particle’s position and velocity to protect sensitive information during subsequent information exchange. In the external learning phase, additional noise is injected into the particle states to protect the learning results. Experimental results demonstrate that the proposed algorithm achieves strong privacy protection while maintaining stable consensus performance, delivering competitive solution quality with low communication cost. This ensures a trade-off between privacy protection and optimization performance.