Scholarlink

<p>Network intrusion detection under realistic operational conditions is complicated by weak flow-level annotation, severe class imbalance, and heterogeneous feature distributions across modern attack campaigns. Multiple Instance Learning (MIL) is well-suited to this setting because only bag-level (incident-level) labels are required. However, no consensus exists on which MIL architectural ingredients are necessary, and on which data regimes they pay off. This paper presents a <b>progressive design study</b> of three novel MIL architectures, all proposed by the authors and presented in monotonically increasing complexity: <b>(i) DG-MIL</b>, a Dual-path Gated MIL baseline with a two-term loss; <b>(ii) HART-MIL</b>, a Hierarchical Adaptive Representation Transformer MIL with Squeeze-and-Excitation gating, dual-path sparse/dense aggregation, and a four-term loss; and <b>(iii) APEX-MIL</b>, an Adaptive Prototype-guided EXpert MIL with a multi-scale Mixture-of-Experts encoder, a Cross-scale Prototype Attention module, an Adaptive Expert Fusion aggregator, and a rigorously-specified five-term training objective. The three architectures are evaluated against five external baselines (AB-MIL, DSMIL, LSTM, 1D-CNN, Random Forest) on three datasets spanning legacy (NSL-KDD), modern (CIC-IDS2018), and contemporary IoT (CIC-IoT2023) regimes. Across 48 main-grid experiments (8 models × 3 datasets × 2 seeds), the <b>Best-of-Proposed</b> — the per-dataset, per-seed winner among DG-MIL, HART-MIL, and APEX-MIL — beats every external baseline on every dataset, with a maximum gain of Δ F1 =  + 0.112 over Random Forest on NSL-KDD. HART-MIL emerges as the modal winner (4 of 6 seeds across the three datasets). A contamination-sensitivity sweep, ANOVA with Tukey HSD multiple-comparison tests, paired t-tests with 95% bootstrap confidence intervals, and a complexity-versus-performance Pareto analysis are reported. The progressive design framing yields an actionable architecture-selection rule for security operations centres: <i>match architectural complexity to dataset heterogeneity rather than maximising it by default.</i></p>

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

Progressive MIL architectures for weakly supervised network intrusion detection

  • 期刊论文

  • 发表日期：

    2026年8月4日

查看全文
  • Sarmad K. Ibrahim,
  • Walaa M. Khalaf,
  • Takele Ferede Agajie

摘要

Network intrusion detection under realistic operational conditions is complicated by weak flow-level annotation, severe class imbalance, and heterogeneous feature distributions across modern attack campaigns. Multiple Instance Learning (MIL) is well-suited to this setting because only bag-level (incident-level) labels are required. However, no consensus exists on which MIL architectural ingredients are necessary, and on which data regimes they pay off. This paper presents a progressive design study of three novel MIL architectures, all proposed by the authors and presented in monotonically increasing complexity: (i) DG-MIL, a Dual-path Gated MIL baseline with a two-term loss; (ii) HART-MIL, a Hierarchical Adaptive Representation Transformer MIL with Squeeze-and-Excitation gating, dual-path sparse/dense aggregation, and a four-term loss; and (iii) APEX-MIL, an Adaptive Prototype-guided EXpert MIL with a multi-scale Mixture-of-Experts encoder, a Cross-scale Prototype Attention module, an Adaptive Expert Fusion aggregator, and a rigorously-specified five-term training objective. The three architectures are evaluated against five external baselines (AB-MIL, DSMIL, LSTM, 1D-CNN, Random Forest) on three datasets spanning legacy (NSL-KDD), modern (CIC-IDS2018), and contemporary IoT (CIC-IoT2023) regimes. Across 48 main-grid experiments (8 models × 3 datasets × 2 seeds), the Best-of-Proposed — the per-dataset, per-seed winner among DG-MIL, HART-MIL, and APEX-MIL — beats every external baseline on every dataset, with a maximum gain of Δ F1 =  + 0.112 over Random Forest on NSL-KDD. HART-MIL emerges as the modal winner (4 of 6 seeds across the three datasets). A contamination-sensitivity sweep, ANOVA with Tukey HSD multiple-comparison tests, paired t-tests with 95% bootstrap confidence intervals, and a complexity-versus-performance Pareto analysis are reported. The progressive design framing yields an actionable architecture-selection rule for security operations centres: match architectural complexity to dataset heterogeneity rather than maximising it by default.