This paper presents a privacy-preserving mechanism for combinatorial double auctions by designing four differential privacy strategies: Laplace-DP, Gaussian-DP, dynamic-DP, and a deep reinforcement learning-based DP. In the proposed system, edge servers and users submit multi-type computing resource supply and demand information, and an auctioneer matches offers based on bid density calculations and honesty indicators to determine transaction prices. To mitigate data leakage risks, noise is injected under the differential privacy framework, with adaptive privacy budget allocation responding to market fluctuations and sensitivity. Experimental results demonstrate that the deep reinforcement learning-based mechanism achieves superior trading success and data utility, while the dynamic approach effectively balances privacy protection and runtime overhead. This work provides novel theoretical insights and practical guidance for optimizing data security and utility in combinatorial double auction.