Sub-headline: BIT researchers introduce Malcom to tackle cross-domain encrypted traffic detection using self-supervised learning.

A significant technical pain point in cybersecurity is the heavy reliance of deep learning models on large-scale labeled datasets for encrypted malware detection. However, acquiring high-quality labels for rapidly evolving malware is costly and time-consuming. When faced with "zero-day" threats or new encryption protocols, traditional models often fail to generalize, leading to a breakdown in defense. This gap between the speed of malware evolution and the slowness of data labeling limits the effectiveness of real-time security monitoring in complex network environments.
In response to these challenges, the research team from Beijing Institute of Technology developed Malcom. This innovation shifts from traditional supervised classification to a self-supervised pre-training paradigm. The architecture leverages a Fully Convolutional Masked Autoencoder (FC-MAE) that randomly masks portions of traffic features. By forcing the network to reconstruct these hidden segments from unlabeled background traffic, Malcom learns robust, high-level representations of data flow. In the fine-tuning stage, the model requires only a tiny fraction of target-domain labels to adapt quickly to new malware variants, effectively "transferring" its generalized knowledge to specific threats.
Research indicates that in experiments on major benchmarks like USTC-TFC, Malcom demonstrates exceptional adaptive performance. Data suggests that even with only 1% of labeled data available, the framework maintains high detection accuracy, significantly surpassing standard CNN and RNN-based methods. Furthermore, the fully convolutional design enhances processing efficiency for high-bandwidth networks. This work provides a reliable technical roadmap for reducing label dependency in network security, offering a robust foundation for building self-evolving and proactive defense systems against encrypted cyber threats.
DOI:10.1007/s11704-025-41273-9