Providing machine unlearning services to users in the cloud, known as Machine Unlearning as a Service (MUaaS), has become a prominent privacy protection strategy. However, existing methods primarily focus on the effectiveness of the unlearning algorithm itself, while neglecting the privacy of the data users upload to the cloud for unlearning. To address this problem, we propose a distribution transformation method that maps users’ private data to a distribution (DT-UPD) that is functionally parallel and orthogonal to the original distribution. We then sample from this transformed distribution and use the sampled data instead of the original data to obtain the unlearning service. In detail, DT-UPD first employs the information bottleneck theory to extract the most relevant features for the unlearning service and then maps these features to a normal distribution. Low-dimensional features are then randomly sampled from this normal distribution and expanded into a high-dimensional space using adversarial reconstruction. This sampling and adversarial reconstruction process removes most of the information from the original private data (including unlearning-related and irrelevant information), ensuring that the final reconstructed data used for the unlearning service has a privacy protection effect. Additionally, we introduce Elastic Weight Consolidation (EWC) loss, which effectively avoids ‘catastrophic unlearning’ during the unlearning process. Extensive experimental results show that DT-UPD can achieve a forgetting accuracy (forget ACC) of 18.22% in the forgetting set with no more than a 6% accuracy loss in the test set.