Adversarial robustness of multi-source data fusion on data text classification supported by information technology

Multi-source heterogeneous text classification faces two major challenges: the difference in feature distribution across different data sources leads to alignment difficulties, and the model is vulnerable to adversarial attacks, resulting in misclassifications. To address this, this paper proposes a multi-source fusion method that integrates hierarchical adaptive BERT with multi-source voting adversarial training. Lightweight adapters are inserted into each Transformer layer of BERT, achieving hierarchical feature projection through learnable affine transformations—lower-level adapters align lexical commonalities, while higher-level adapters capture semantic specificities, and a gating mechanism dynamically fuses multi-source representations. Furthermore, a gradient-symmetric cross-source perturbation generator is constructed, retaining only perturbations that enable at least two source models to predict divergence, generating transferable adversarial examples. During training, a multi-source classifier committee and hard voting mechanism are employed: when at least 50% of the sub-models detect anomalies, adversarial loss is triggered, forcing the model to learn cross-source robust features. Simultaneously, the fusion weights are dynamically adjusted based on the confidence of each source, and feature dimensionality reduction is performed on low-confidence sources to suppress noise. Experiments show that this method performs well in feature alignment (mean cosine similarity 0.842, HSIC 0.615, low standard deviation) and adversarial robustness (accuracy 87.8%, F1 85.4%). After introducing multi-source voting, the false negative rate (FNR) is 0.12–0.18 and the false positive rate (FPR) is 0.05–0.08 under five types of attacks: FGSM, PGD, TextBugger, DeepWordBug, and TextFooler, demonstrating good cross-source adversarial detection capabilities.