This paper studies the problem of achieving group fairness in the practical setting where the sensitive attribute of a tiny portion of the training data is accessible. Existing fairness-aware algorithms typically train a sensitive attribute classifier from the subset of training data with sensitive attribute to impute the training data with missing sensitive attribute. Unfortunately, this direct approach has been proven ineffective because the aforementioned classifier will produce erroneous pseudo-labels for sensitive attributes, which will mislead the model’s fair training. In this paper, we propose a Pseudo-Label Error Aware (PLEA) framework that can assist existing fairness-aware learning algorithms in achieving group fairness. Specifically, PLEA also uses an additional classifier to predict pseudo-labels for sensitive attributes. However, this framework mitigates the negative impact of erroneous pseudo-labels on the model’s fair training, i.e., its fairness performance, through two approaches: 1) resampling based on sample weights and 2) computing weighted fairness loss in bins, where the sample weights are inversely proportional to the error of the sensitive attribute pseudo-labels. Our proposed framework enhances existing fairness-aware learning algorithms by leveraging samples that lack sensitive labels, aiming to improve their fairness performances in such real-world application scenarios. Experiments on two real-world datasets demonstrate that PLEA can significantly enhance the fairness performance, with almost no loss in accuracy.