In order to address the issues of low-quality candidate sets, severe noise accumulation, and insufficient frequency-estimation accuracy in existing local differential privacy–based frequent itemset mining methods for high-dimensional sparse data, a two-stage adaptive perturbation algorithm incorporating a domain-adaptive perturbation strategy was used to investigate high-accuracy frequent itemset mining mechanisms under high-dimensional sparsity. The results show that initial candidate noise is effectively reduced and prefix-based filtering quality is improved by dynamically selecting perturbation mechanisms in the candidate-generation stage; local variance is significantly suppressed and global frequency reconstruction is optimized by adaptively choosing mechanisms according to subdomain size in the frequency-estimation stage; and superior performance of the proposed algorithm over existing methods in estimation accuracy, frequent-item discovery ability, and runtime is demonstrated by experiments on four real-world datasets, with up to a 15% improvement in Precision and Recall achieved on highly sparse datasets. It is concluded that frequent itemset mining performance on high-dimensional sparse data can be substantially enhanced by the proposed two-stage adaptive perturbation algorithm while preserving local differential privacy.