Scholarlink

Modern network environments demand intrusion detection systems capable of processing heterogeneous data while quantifying uncertainty and maintaining adversarial robustness. This paper introduces a Stochastic Multimodal Transformer architecture that integrates three core innovations: (1) stochastic attention mechanisms with Gaussian noise injection for epistemic uncertainty modeling, (2) specialized multimodal encoders (CNN/LSTM/GRU) for traffic, logs, and API traces with Gaussian Process uncertainty quantification, and (3) comprehensive adversarial training for enhanced robustness. Our architecture processes diverse network modalities through specialized encoders, fuses them via stochastic transformers with uncertainty injection, and provides calibrated confidence estimates through sparse Gaussian Process layers. Comprehensive evaluation on three benchmark datasets (CIC-IoT-2023, CSE-CICIDS2018, UNSW-TON-NB2015) demonstrates superior performance with 98.3% average accuracy, exceptional adversarial robustness (94.7% under GAN attacks), and well-calibrated uncertainty estimates (ECE = 0.031). The framework maintains only 3.6% performance degradation under sophisticated adversarial scenarios while achieving 4.8% improvement over state-of-the-art baselines.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Stochastic Multimodal Transformer with Uncertainty Quantification for Robust Network Intrusion Detection

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Roger Nick Anaedevha,
  • Alexander Gennadievich Trofimov

摘要

Modern network environments demand intrusion detection systems capable of processing heterogeneous data while quantifying uncertainty and maintaining adversarial robustness. This paper introduces a Stochastic Multimodal Transformer architecture that integrates three core innovations: (1) stochastic attention mechanisms with Gaussian noise injection for epistemic uncertainty modeling, (2) specialized multimodal encoders (CNN/LSTM/GRU) for traffic, logs, and API traces with Gaussian Process uncertainty quantification, and (3) comprehensive adversarial training for enhanced robustness. Our architecture processes diverse network modalities through specialized encoders, fuses them via stochastic transformers with uncertainty injection, and provides calibrated confidence estimates through sparse Gaussian Process layers. Comprehensive evaluation on three benchmark datasets (CIC-IoT-2023, CSE-CICIDS2018, UNSW-TON-NB2015) demonstrates superior performance with 98.3% average accuracy, exceptional adversarial robustness (94.7% under GAN attacks), and well-calibrated uncertainty estimates (ECE = 0.031). The framework maintains only 3.6% performance degradation under sophisticated adversarial scenarios while achieving 4.8% improvement over state-of-the-art baselines.