Adaptive Pruning of Deep Neural Networks for Resource-Aware Embedded Intrusion Detection on the Edge

Alexandre Broggi, Nathaniel D. Bastian, Lance Fiondella, Gökhan KularXiv 2025 · DOI · arXiv

Introduction

Deep Neural Networks (DNNs) have demonstrated exceptional efficacy in identifying cyberattacks within network traffic. However, these models are typically computationally intensive, requiring significant memory and processing power. This poses a substantial challenge for the Internet of Battlefield Things (IoBT) and edge computing environments, where security solutions must operate on resource-constrained, low-power hardware.

Model pruning—the process of removing redundant neurons or weights—is a standard technique for compressing DNNs. While extensively studied in computer vision (CNNs) and natural language processing (Transformers), its application to Network Intrusion Detection Systems (NIDS) remains under-explored. NIDS often rely on simpler, fully connected architectures and tabular data, which may respond differently to pruning "tricks" optimized for spatial or sequential data. This paper evaluates eight pruning strategies and introduces a novel method, Iterative Theseus, specifically designed to enhance efficiency in these contexts.

Methodology

The researchers evaluated several pruning paradigms on a deep fully connected network architecture. These included:

  • Baselines: Random Pruning and "Recreation" (training a smaller architecture from scratch).
  • State-of-the-Art: ADMM-joint, DAIS (Differentiable Annealing Indicator Search), and ThiNet.
  • Knowledge Distillation: Bert-Theseus, which replaces network layers with compressed versions.
  • Iterative Theseus: A novel evolution of Bert-Theseus that replaces and retrains layers iteratively rather than in large blocks, allowing for more granular control over the compression process.
Comparison of unpruned, unstructured, structured, and hybrid pruning techniques.

The study categorized pruning into structured (removing entire filters or neurons), unstructured (removing individual weights), and hybrid approaches. Structured pruning is generally more hardware-friendly as it results in dense matrices of smaller dimensions, whereas unstructured pruning requires specialized hardware to exploit sparsity.

Experimental Setup

The evaluation utilized the ACI-IOT-2023-payload dataset, which captures real-world IoT network traffic. The dataset includes various attack vectors such as DNS Floods, Slowloris, SYN Floods, and Port Scans. Raw packet data was preprocessed into 1,515 feature columns. To address the inherent class imbalance in cybersecurity data (where benign traffic significantly outweighs attack traffic), the authors employed under-sampling and class grouping techniques.

Results and Discussion

The experiments revealed a distinct split between "Helped" and "Hindered" algorithms. Methods like DAIS and Bert-Theseus occasionally improved accuracy as the model size decreased, suggesting that the initial models were over-parameterized for the specific NIDS task. Conversely, "Hindered" algorithms experienced performance degradation, often reaching a "critical point" where the removal of essential nodes caused the F1 score to plummet.

Normalized F1 score vs. number of non-zero parameters for various pruning methods.

A significant finding was that high-performance pruning algorithms designed for CNNs, such as ADMM, did not always translate effectively to the fully connected layers of NIDS. This underscores the need for architecture-specific optimization.

Iterative Theseus emerged as the top performer, achieving the highest F1 scores (up to 0.971) across most pruning percentages. However, it was also the most computationally expensive to execute. For practical real-world deployment where pruning time is a factor, ThiNet was identified as the best compromise, offering high performance with significantly faster execution.

Conclusion

This research demonstrates that edge-based NIDS is highly feasible, with models often compressible to 12%–25% of their original size without significant loss in accuracy. While Iterative Theseus provides superior compression quality, the choice of pruning method must balance accuracy with the computational cost of the pruning process itself. Future work should focus on developing pruning techniques specifically tailored for the non-spatial, tabular data characteristic of network logs to further optimize security on the edge.

Summary updated

BibTeX
@misc{broggi2025adaptive,
  author = {Alexandre Broggi and Nathaniel D. Bastian and Lance Fiondella and Gökhan Kul},
  title = {{Adaptive Pruning of Deep Neural Networks for Resource-Aware Embedded Intrusion Detection on the Edge}},
  year = {2025},
  eprint = {2505.14592},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.14592},
}
Figure