A Novel Internet of Things-Centric Framework to Mine Malicious Frequent Patterns

Nighat Usman, Qaisar Javaid, Adnan Akhunzada, Kim-Kwang Raymond Choo, Saeeda Usman, Asma Sher, Manzoor Ilahi, Masoom AlamIEEE Access 2019 · DOI · OpenReview

Introduction

The rapid expansion of the Internet of Things (IoT) has introduced a vast array of connected devices, from smart home appliances to critical industrial sensors. While these devices enhance connectivity, they also present a significant attack surface for cyber threats. Detecting "malicious frequent patterns"—sequences of activities indicative of an ongoing attack—is a cornerstone of IoT defense. However, traditional Association Rule Mining (ARM) algorithms, such as Apriori and FP-Growth, are often too computationally intensive or memory-heavy for the resource-constrained environments typical of IoT hardware.

Methodology: The MMFP Framework

To address these limitations, researchers have developed the Mine Malicious Frequent Patterns (MMFP) framework. This lightweight approach streamlines the data mining process through a three-phase methodology designed for efficiency and accuracy.

Feature Selection and Preprocessing

The first phase utilizes Principal Component Analysis (PCA) to identify the most significant features within security logs, such as IP addresses or packet sizes. By discarding redundant information, the framework reduces the initial data volume.

The second phase introduces a novel "masking" technique. As illustrated in the flowchart below, a user-defined mask assigns values (0, 1, or 2) to features, determining whether they are ignored, selected as-is, or converted into numerical "legends."

The MMFP preprocessing framework utilizing a user-defined mask to optimize dataset features.

This masking process is critical for normalizing diverse data types into a unified numerical format, significantly lowering the input/output (I/O) costs and memory overhead during analysis. By converting complex string data into legends, the dataset becomes much smaller and easier for low-power IoT gateways to process.

The MMFP Algorithm

The core MMFP algorithm is engineered to scan datasets efficiently. Unlike traditional methods that may require multiple passes or build complex tree structures, MMFP discards non-frequent patterns early in the process. This ensures that the generated patterns are "exact," avoiding the artificial gaps often found in other sequential mining techniques like Prefixspan.

Experimental Results

The framework was rigorously tested using a real-world dataset of over one million security records provided by Trillium Pakistan. Comparative analysis against established algorithms demonstrated MMFP's superior performance in resource-constrained scenarios:

  • Memory Efficiency: MMFP utilized only 9.05 MB of memory, a stark contrast to the 202.25 MB required by FP-Growth.
  • Processing Speed: The task completion time for MMFP was 63 ms, significantly faster than Prefixspan (419 ms) and FP-Growth (1833 ms).
  • Pattern Quality: MMFP produced 174 concise and relevant patterns, avoiding the "pattern explosion" seen in FP-Growth, which generated over 921,000 itemsets, many of which were redundant noise.
  • Rule Generation: MMFP proved to be nearly twice as fast as the Apriori algorithm in generating the final association rules used for anomaly detection.

Conclusion and Future Work

The MMFP framework offers a robust, lightweight solution for IoT security, balancing high detection accuracy with minimal resource consumption. While currently tested on specific organizational datasets, future research will focus on scaling the framework for diverse network architectures and integrating deep learning components to enhance automated threat response systems and intra-device communication analysis.

Summary updated

Abstract

There are a number of research challenges associated with Internet of Things (IoT) security, and one of these challenges is to design novel frameworks to mine malicious frequent patterns for identifying misuse and detecting anomalies without incurring high computational costs (e.g., due to generation and analysis of unnecessary patterns and gap creation between patterns). Association rule mining is a popular approach in the literature; hence, in this paper, we critically analyze existing association rule mining techniques. We then present a framework for mining malicious frequent patterns in an IoT deployment, prior to evaluating the utility of the proposed framework using data from a Pakistan-based organization.
BibTeX
@article{usman2019novel,
  author = {Nighat Usman and Qaisar Javaid and Adnan Akhunzada and Kim-Kwang Raymond Choo and Saeeda Usman and Asma Sher and Manzoor Ilahi and Masoom Alam},
  title = {{A Novel Internet of Things-Centric Framework to Mine Malicious Frequent Patterns}},
  journal = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2017.2690456},
}
Figure