Improving the effectiveness of APT attack detection models is one of the most critical and essential tasks today. Following this trend, this paper proposes a new model called ACDF-mLSTM to address two primary challenges currently faced by research in this field: (i) data imbalance and (ii) information aggregation and feature extraction. Specifically, to solve the data imbalance problem, the paper proposes a novel data generation method named ACDF. This method leverages advanced and sophisticated techniques to focus on identifying crucial points in sequential data and analyzing context by considering preceding and succeeding data points. Subsequently, a Diffusion Model is applied to generate synthetic APT attack data, built upon the principle of gradual diffusion. With this approach, the ACDF model can generate more meaningful and realistic data. Next, to address the task of information aggregation and feature extraction, the paper proposes a new deep learning model named mLSTM, based on the optimization of Long Short-Term Memory (LSTM). Thus, the mLSTM model performs two main tasks: (i) extracting information from network flows within traffic and (ii) aggregating and highlighting important information before it enters the classification model. In the experimental section, the paper evaluates the ACDF-mLSTM model for the first time across various scenarios and datasets to demonstrate its effectiveness and adaptability. The evaluation results show that the ACDF-mLSTM model outperformed most other methods by an average of 2 to 12% across all metrics and on all experimental datasets.