A Survey of Machine Learning-Based Zero-Day Attack Detection: Challenges and Future Directions.

Area of Science:

• Cybersecurity
• Computer Science
• Artificial Intelligence

Background:

• Zero-day attacks pose a significant threat by exploiting unknown software vulnerabilities.
• Traditional signature-based detection is ineffective against zero-day threats due to the lack of prior signatures.
• Machine learning (ML) presents a viable alternative for detecting zero-day attacks by analyzing behavioral patterns.

Purpose of the Study:

• To conduct a comprehensive survey of ML-based approaches for zero-day attack detection.
• To compare ML models, datasets, and evaluation metrics used in existing studies.
• To identify challenges and recommend future research directions in ML-based zero-day attack detection.

Main Methods:

• Systematic literature review of ML-based zero-day attack detection techniques.
• Comparative analysis of different ML algorithms, including their application and performance.
• Evaluation of datasets used for training and testing ML models.
• Identification of key performance indicators such as accuracy, recall, and uniformity.

Main Results:

• Existing ML-based methods show promise but exhibit limitations in accuracy, recall, and uniformity.
• A wide range of ML models have been applied, with varying degrees of success.
• Significant heterogeneity exists in the datasets and evaluation methodologies across studies.

Conclusions:

• Despite advancements, current ML-based zero-day attack detection methods require improvement to address accuracy and consistency issues.
• Further research is needed to overcome the identified challenges and enhance the robustness of ML solutions.
• Future work should focus on developing more uniform and accurate detection strategies for diverse zero-day attack vectors.