Computer Network Virus Defense with Data Mining-based Active Protection

Xiaohong Li, Yang Li, Hong HeScalable 2024 · DOI

TL;DR

This paper introduces DMAV (Data Mining-based Antivirus), a proactive defense system that uses Object-Oriented Analysis (OOA) mining to identify malware by its "behavioral DNA"—specifically, sequences of Windows API calls. By moving beyond static signatures, the system achieves a 92% recognition rate for unknown viruses while significantly reducing scanning time and false positives compared to traditional industry tools.

What Problem Does This Paper Address?

The cybersecurity landscape is a perpetual arms race. Traditional antivirus software relies heavily on signature-based detection, which works like a "Most Wanted" poster: if a file's code matches a known criminal's profile, it’s blocked. However, hackers have evolved "metamorphic" or deformed viruses that change their internal structure with every infection, rendering static signatures useless.

Furthermore, zero-day (unknown) threats pose a massive risk because no signature exists for them yet. While newer techniques like sandboxing and heuristic analysis exist, they often suffer from high computational overhead (slowing down the system) or high false-alarm rates. The authors aim to bridge this gap with a system that is both intelligent enough to spot new threats and efficient enough for real-time protection.

Methods

The authors propose the Data Mining-based Antivirus (DMAV) system, which consists of three core modules:

  1. PE File Parser: When a suspicious Portable Executable (PE) file is scanned, this module extracts the sequence of Windows API calls from its import table. These calls (e.g., opening a file, connecting to a network) represent the actual actions the program intends to take. Each API is mapped to a unique 32-bit integer ID to create a "feature vector."
  2. Rule Generator (OOA Mining): Using an Object-Oriented Analysis mining approach based on the Apriori algorithm, the system analyzes a large database of known malware to find frequent patterns. It generates association rules that link specific API sequences to malicious categories like Trojans, Worms, or Spyware. This helps the system learn the "intent" behind certain code patterns.
  3. Similarity Measurement: To detect deformed viruses, the system doesn't look for an exact match. It uses a sequence rearrangement algorithm (represented by a matrix) to align the API sequence of a suspicious file with known malicious rules. It then calculates similarity using multiple metrics, including Cosine similarity, Jaccard extended cosine, and Pearson correlation, to determine if the file is a threat.

Datasets

The researchers utilized a dataset sourced from the Jinshan Antivirus Laboratory:

  • Training Set: 5,000 virus samples and 2,000 non-virus (benign) samples used to extract features and generate the OOA rules.
  • Test Set: 150 virus samples (including deformed versions of well-known malware like Beagle, Blaster, and Mydoom) and 500 non-virus samples.

Key Insights and Ideas

  • API Calls as Behavioral Fingerprints: The core "aha moment" is that while a virus can change its code (obfuscation), it cannot easily change the underlying system calls it needs to perform its malicious task. The sequence of these calls is a much more robust indicator of intent than the raw code.
  • Overcoming "Over-learning": The authors argue that common machine learning models like decision trees often "over-learn" (overfit) on training data, making them brittle against new threats. OOA mining provides a more generalized set of rules that better captures the essence of malicious behavior.
  • Matrix-Based Rearrangement: By using a matrix to rearrange sequences before calculating similarity, the system can effectively ignore "noise" or "padding" instructions that malware authors insert to confuse scanners.

Key Results

The DMAV system was benchmarked against several industry-standard and open-source scanners (referred to in the paper as N, M, D, K, and SAVE):

  • Superior Accuracy: DMAV successfully identified all deformed versions of the test viruses, whereas most commercial scanners missed several variants (e.g., Beagle3 and Blaster2).
  • High Detection Rate: It achieved a 92% recognition rate for unknown viruses, significantly higher than the 32%–54% range seen in the comparison group.
  • Efficiency: DMAV was the fastest, with a scanning time of 190ms, compared to 250ms–450ms for other systems.
  • Low False Positives: In a test of 500 benign files, DMAV only flagged 2 incorrectly, while other systems flagged between 4 and 15.

Limitations and Future Work

  • Hardware Context: The experiments were conducted on an Intel P4 1GHz processor with 1GB of RAM. While this demonstrates the efficiency of the algorithm on low-end hardware, modern malware often employs more sophisticated anti-analysis techniques that might require more modern benchmarking.
  • Static Analysis Limits: The current parser uses static analysis (reading the file without running it). Sophisticated malware that uses dynamic API loading or heavy encryption might bypass this initial scan.
  • Future Directions: The authors suggest integrating deep learning for even more precise behavioral models and moving toward multimodal data analysis—combining API sequences with text, image, and network traffic data to create a truly holistic defense system.

Summary updated

Abstract

A novel approach is presented in this paper to address the limitations of virtual machine technology, active kernel technology, heuristic killing technology, and behaviour killing technology in computer network virus defence. The proposed method provides data mining technology, specifically Object-Oriented Analysis (OOA) mining, to detect deformed and unknown viruses by analyzing the sequence of Win API calls in PE files. Experimental results showcase the Data Mining-based Antivirus (DMAV) system's superiority over existing virus scanning software in multiple aspects: higher accuracy in deformed virus detection, enhanced active defence capabilities against unknown viruses (with a recognition rate of 92%), improved efficiency, and a reduced false alarm rate for non-virus file detection. Furthermore, the paper introduces an OOA rule generator to optimize feature extraction, enhancing the system's intelligence and robustness. This research provides a promising solution to support virus detection accuracy, active defence mechanisms, and overall efficiency while minimizing false positives in virus scanning, thus contributing significantly to the advancement of computer network security.
BibTeX
@inproceedings{li2024computer,
  author = {Xiaohong Li and Yang Li and Hong He},
  title = {{Computer Network Virus Defense with Data Mining-based Active Protection}},
  booktitle = {Scalable},
  year = {2024},
  doi = {10.12694/scpe.v25i1.2173},
}
Figure