MEFA: Multisource Entropy‐Weighted Feature Adaptation for Cross‐Domain Intrusion Detection

With the development of network technologies, security problems are becoming more and more serious. Recently, intrusion detection systems (IDS) have been considered an effective method for identifying and preventing malicious activities; however, due to the lack of labeled data in new environments, it is challenging to train reliable models. Moreover, the differences in traffic patterns and feature distributions across domains make knowledge transfer more difficult. To address this issue, we propose a novel framework named MEFA (Multisource Entropy‐weighted Feature Adaptation), which helps train an IDS in a new target domain by leveraging knowledge from multiple source domains. The proposed method comprises three key components. First, the target domain data are compressed into a latent feature space so that their representations can be unified with those of source domain models. Second, a semantic‐guided discriminative alignment is introduced between the target and each source domain to improve feature consistency and enhance alignment quality. Third, we introduce an entropy‐based pseudo‐label learning that assigns adaptive weights to each source model based on its prediction confidence on target samples. The fused soft labels are then used to train the final target classifier. Experiments on four public intrusion detection datasets demonstrate that MEFA consistently outperforms existing domain adaptation approaches under both unsupervised and few‐shot settings, which demonstrates an average classification accuracy of 89.17% across all four target domains.