Leveraging Semi-Supervised Learning to Reduce Labeled Data Requirements in Intrusion Detection
Simone Albero, Tommaso Caiazzi, Stefano Iannucci, Paolo Merialdo, Riccardo Torlone · 2025
Deep learning-based intrusion detection systems often depend on large labeled datasets and generating such data is both costly and sometimes impractical. To overcome this limitation, we propose a hybrid learning approach built on a transformer architecture. Our method integrates a self-supervised pretraining phase, where the model is trained to reconstruct noised segments of input traffic data from unlabeled sequences, with a supervised fine-tuning stage that requires only a fraction of labeled data. Our experiments demonstrate the effectiveness of this hybrid approach, achieving up to 98.8% of the performance of the supervised models using 50% of the labeled data. Index Terms—Intrusion Detection, Hybrid Learning