Membership inference attacks (MIAs) have emerged as a prominent focus in machine learning security, seeking to determine whether a particular data point was used during a model’s training. Previous attack methods leveraged the outputs of the target model to distinguish between member and non-member samples. However, when the model outputs for these two types of samples are similar, such methods often fail to make accurate distinctions. Recent difficulty calibration methods and their improvements aim to mitigate this issue. However, since they rely solely on the outputs of the final model state for calibration, they still misclassify some non-members, leading to an elevated rate of false positives. To overcome this limitation, we develop a new attack approach termed LMTID, which Leverages Multiple metrics obtained during the model Training to Improve Difficulty calibration. More specifically, LMTID utilizes knowledge distillation to produce multiple student models, each corresponding to a distinct phase in the target model’s training process. Subsequently, for each target sample, we compute multiple metrics on the distilled models to construct a multi-metric sequence. This sequence is concatenated with the calibrated membership score and used as input to the attack model to determine whether the sample is a member. Extensive experiments on five datasets and five model architectures demonstrate that LMTID outperforms all baselines. In particular, on TPR@0.1%FPR, LMTID attains performance up to 2.5 times higher than the current state-of-the-art, highlighting its substantially enhanced attack effectiveness.