Abstract
Cyber-attack prevention in online retailing has proven to be a challenging task. Machine learning (ML) algorithms play a significant role in preventing cyber-attacks. However, existing studies on ML-based prevention techniques are not well-researched. To bridge the gap in the literature, this article reviews existing literature on cybercrime prevention using ML techniques in an online retail context. A systematic literature review (SLR) was conducted of the literature on ML-driven cyber-attack prevention techniques in e-tailing, starting with 1828 relevant publications from four databases using the PRISMA approach. Finally, fifty-four journal articles from 2018 to 2023 were analysed. The review revealed that the research on ML prevention algorithms in e-tailing is an emerging field with a growing number of articles in recent years, and significant emphasis has been placed on supervised and unsupervised methods, with a particular focus on classification techniques, e.g., support vector machine and naive Bayes for prevention of cybercrimes in e-tailing. The SLR highlights several technical problems and offers suggestions for further study. Our research extends existing knowledge by synthesising existing literature and highlights cutting-edge findings and ML methods on cyber-attack prevention in online retailing. It also presents a holistic view of past research and offers the foundation for future studies. Furthermore, this SLR will help online retailers in knowledge provision and improve their ability to develop and implement ML practices to prevent cybercrimes.
Introduction
Online retail (e-tailing) has become a primary target for cybercriminals, with reports suggesting that retailers experience the highest number of data breaches of any sector. While machine learning is a known powerhouse for cybersecurity, existing research remains fragmented. Most studies focus on general "detection" rather than proactive "prevention," and few offer a cohesive strategy specifically for the retail environment.
Retailers face a unique "prevention dilemma": they must handle massive volumes of sensitive customer data and financial transactions in real-time while defending against increasingly sophisticated threats like phishing, DDoS attacks, and "insider" fraud. This paper aims to bridge the gap between theoretical ML capabilities and practical, robust implementation strategies for the retail industry.
Methodology
The authors conducted a Systematic Literature Review (SLR) following the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines to ensure a rigorous and reproducible selection process.
- Search Strategy: Four major databases (Scopus, Emerald Insight, Science Direct, and IEEE Explore) were queried using keywords related to cyber-attacks, online retail, and ML.
- Filtering Process: Starting with an initial pool of 1,828 publications, the authors applied strict inclusion and exclusion criteria to narrow the field down to 54 high-impact, peer-reviewed journal articles published between 2018 and 2023.
The Integrative Framework
A core contribution of this study is the application of Situational Crime Prevention (SCP) theory—a framework traditionally used in physical security—to the digital retail space. By integrating SCP with machine learning paradigms, the authors propose a model where ML acts as an autonomous agent to "harden the target."
As illustrated in the framework, the system processes diverse data inputs—including customer profiles, financial transactions, network traffic, and historical attack logs—through three primary ML paradigms:
- Supervised Learning (SL): Used for known threat patterns.
- Unsupervised Learning (UL): Used for anomaly detection and zero-day threats.
- Reinforcement Learning (RL): Used for adaptive strategies against evolving attacks.
Key Insights and Analysis
The Shift to Proactive Prevention
The paper emphasizes that ML must move beyond identifying a breach after it happens. Prevention requires "implanting" ML like a "secret agent" that supports data protection autonomously. This involves reducing provocations and removing excuses for criminal behavior by increasing the perceived effort and risk for the attacker.
Dominance of Classification
Supervised learning remains the most frequently applied method in the reviewed literature. Specifically, Support Vector Machines (SVM) and Naive Bayes are identified as the "workhorses" for detecting phishing and fraudulent transactions due to their high accuracy with labeled data.
The Role of Deep Learning and Hybrid Models
For complex, non-linear threats, Deep Learning models (CNN, RNN, LSTM) are becoming essential. They are particularly effective at identifying "Advanced Persistent Threats" (APTs) that might go undetected by simpler algorithms. Furthermore, the review highlights a growing trend toward hybrid models—combining supervised and unsupervised methods—to catch both known attack signatures and previously unknown anomalies.
Results and Future Directions
The authors noted a significant upsurge in publications from 2018 to 2023, reflecting an escalating "arms race" between retailers and cybercriminals. While the technical capabilities of ML are expanding, the review identifies several critical gaps for future research:
- Empirical Diversity: There is a need for more studies comparing multiple ML frameworks in real-world retail settings rather than in isolation.
- Human Factors: Most literature focuses on the technical "black box," often ignoring the human-in-the-loop aspect of security and organizational culture.
- Real-time Efficiency: High latency in an ML prevention system could negatively impact the customer experience. Future research must balance computational cost with real-time processing needs.
In conclusion, by synthesizing existing literature through the lens of SCP, this study provides a holistic foundation for retailers to develop more robust, adaptive, and efficient cyber-attack prevention strategies.
Summary updated