Abstract
At present, machine learning, especially active learning in machine learning, is introduced in the cyberspace security field. In this paper, we first analyze the background of cyberspace security and profile active learning. We then focus on a survey of the application research of active learning in the cyberspace security field, mainly including system software security, network security, and application security. Finally, we explore potential future issues regarding active learning in cyberspace security.
Introduction
The fundamental challenge in modern cybersecurity is the sheer volume and complexity of data. While Machine Learning (ML) offers a way to automate threat detection, most effective models rely on supervised learning, which requires massive amounts of labeled data. Labeling security data is incredibly "expensive" in terms of time and expertise, requiring highly skilled security analysts to manually inspect code or logs. As threats evolve rapidly, the time it takes to label a dataset often means the resulting model is outdated by the time it is deployed. Active Learning (AL) addresses this "labeling bottleneck" by intelligently selecting only the most informative data points for human experts to review.
The Active Learning Framework
Active Learning shifts the paradigm from passive data processing to an interactive, "human-in-the-loop" approach. The framework typically consists of a learning engine (the classifier) and a selection engine (the algorithm that chooses which data to label).
As illustrated in the figure above, the process is iterative. The machine learning model (C) queries samples from an unlabeled pool (U). These selected queries are presented to a supervisor (S), typically a human expert, who provides the correct labels. These newly labeled samples are added to the training set (L), and the model is retrained, improving its accuracy with each cycle.
Core Selection Strategies
Researchers in the security domain primarily utilize three selection strategies:
- Uncertainty Sampling: The most common approach, where the model queries the samples it is least confident about—specifically those located near the decision boundary.
- Query-By-Committee (QBC): Multiple models "vote" on a sample; the samples with the most disagreement are sent to the expert for labeling.
- Expected Error/Variance Reduction: Advanced statistical methods that choose samples likely to minimize the model's future error or overall variance.
Domain Applications in Cybersecurity
The survey details how these methods are applied across three specific domains:
- System Software Security: This includes vulnerability mining and malicious code analysis. For instance, the HARMLESS tool achieved 90-99% recall while requiring experts to review only 24-45% of the source code.
- Network Security: AL is used for BGP anomaly detection, botnet identification, and Intrusion Detection Systems (IDS). In one study, a model achieved 90% classification accuracy by having a human label only 0.13% of 4.9 million records.
- Application Security: This covers spam filtering, malicious URL recognition, and PDF-based malware detection. A cost-sensitive online AL approach (CSOAL) outperformed standard models by querying only 0.5% of a 1-million-instance dataset.
Key Insights and Results
One of the most striking insights from the research is the scale of efficiency. By focusing on "uncertain" data—the samples that fall between obviously malicious and obviously benign—models learn the boundary of what constitutes a threat much faster. This leads to a drastic reduction in labor, with some frameworks reducing detection error by 36.5% while using only 1% of the total available samples for training.
Furthermore, AL is naturally suited for "Zero-Day" readiness. Because it focuses on uncertainty, it can identify new threats that do not match existing signatures but appear "confusing" to a standard classifier, prompting immediate human intervention.
Challenges and Future Directions
Despite its promise, several hurdles remain for the widespread adoption of AL in security:
- The Dimensionality Disaster: Security data often features hundreds of dimensions. Calculating which samples are most informative in high-dimensional space is computationally expensive.
- Domain Knowledge Integration: Current AL algorithms often operate as "black boxes." Future research must better incorporate a priori expert knowledge into the selection engine to make queries more efficient.
- Unbalanced Data: In real-world scenarios, malicious samples are rare (the "needle in a haystack" problem). Future AL models must be more cost-sensitive to ensure they do not miss rare, critical threats while becoming proficient at identifying benign traffic.
Conclusion
Active Learning represents a significant step forward in making machine learning practical for the fast-paced world of cyberspace security. By reducing the manual labeling burden by up to 99% in some cases, it allows security teams to keep pace with evolving threats while maintaining high detection accuracy.
Summary updated