Membership inference attack is a privacy attack against machine learning models. It aims to ascertain whether a particular data sample belongs to the model’s training set. Existing membership inference attack methods rely heavily on overfitting. However, the outputs of some members and non-members are comparable in reality. Although some current work attempts to tackle this problem, recognizing members accurately is still a challenging task. To address the above problem, this paper proposes a new attack algorithm named CFN-MIA. CFN-MIA utilizes the Jacobian matrix to quantify the prediction variations between members and non-members. In addition, this method incorporates difficulty calibration to provide clearer membership signals through multi-feature learning. In this paper, the algorithm is evaluated on multiple datasets and compared to three different MIAs. The results illustrate that CFN-MIA can significantly improve the attack performance and outperform the other three inference attacks.