| 摘 要: 针对入侵检测及相关安全分类任务中高维特征、冗余信息与复杂非线性模式识别的挑战,本文提出一种混合感知深度神经网络模型 HyPerModE。该模型融合了仿生嗅觉侧抑制、频谱域特征调制与残差门控多分支融合机制,旨在提升表格型安全数据的特征表征能力与泛化稳定性。具体而言,模型首先通过特征变换器将原始表格特征映射为统一上下文嵌入；随后并行引入嗅觉侧抑制分支、频谱滤波分支与时域残差分支,从稀疏抑制、全局频域依赖和非线性残差建模三个角度提取互补表征；最后采用以全局投影为主干的残差门控融合模块,自适应调节不同感知分支对最终表示的贡献。实验在 NSL-KDD、Phishing-Websites 和 Spambase 三个安全相关数据集上进行。结果表明,在严格无数据泄露的评估设置下,HyPerModE-Full 在 NSL-KDD 上取得 0.8119 的 macro-F1,明显优于随机森林的 0.7868,并接近 MLP 的 0.8203；在 Phishing-Websites 上达到 0.9738 的 macro-F1,与 MLP 持平并接近随机森林；在 Spambase 上取得 0.9376 的 macro-F1,略高于 MLP 但低于随机森林。消融实验表明,不同感知分支的贡献具有明显数据集依赖性,残差门控融合能够在多任务场景下维持稳定的强基线级性能。本文为安全表格数据建模提供了一种融合生物启发与频谱感知机制的深度网络设计思路。 |
| 关键词: 入侵检测,仿生嗅觉,频谱感知,全息融合 |
|
中图分类号:
文献标识码:
|
| 基金项目: 天水市自然科学基金2025-FZGHK-4856 |
|
| A multimodal hybrid perception deep network model for intrusion detection |
|
hanyin1, weikaibin2, xiexiannian2
|
1.Tianshui Productivity Promotion Center;2.Tianshui Normal University
|
| Abstract: To address the challenges of high-dimensional features, redundant information, and complex nonlinear pattern recognition in intrusion detection and related security classification tasks, this paper proposes a hybrid perception deep neural network model, HyPerModE. This model integrates bionic olfactory lateral inhibition, spectral domain feature modulation, and residual gated multi-branch fusion mechanisms, aiming to enhance the feature representation ability and generalization stability of tabular security data. Specifically, the model first maps the original tabular features to a unified context embedding through a feature transformer; then, it introduces olfactory lateral inhibition, spectral filtering, and time-domain residual branches in parallel, extracting complementary representations from the perspectives of sparse inhibition, global spectral domain dependency, and nonlinear residual modeling; finally, it adopts a residual gated fusion module with global projection as the backbone to adaptively adjust the contribution of different perception branches to the final representation. Experiments were conducted on three security-related datasets: NSL-KDD, Phishing-Websites, and Spambase. The results show that under a strict no-data-leakage evaluation setting, HyPerModE-Full achieves a macro-F1 of 0.8119 on NSL-KDD, significantly outperforming random forest''s 0.7868 and approaching MLP''s 0.8203; it reaches a macro-F1 of 0.9738 on Phishing-Websites, on par with MLP and close to random forest; and it achieves a macro-F1 of 0.9376 on Spambase, slightly higher than MLP but lower than random forest. Ablation experiments indicate that the contribution of different perception branches has significant dataset dependence, and residual gated fusion can maintain stable strong baseline-level performance in multi-task scenarios. This paper provides a design idea for deep network models that integrate bio-inspired and spectral perception mechanisms for security tabular data modeling. |
| Keywords: Intrusion detection, bionic olfaction, spectrum sensing, holographic fusion |
