Designing a Dynamic Risk Assessment Framework for Machine Learning-Specific Threats in Enterprise MLSecOps Implementation

Articles in Press

Document Type : Original Article

Authors
Hoodad Tabibi 1
Elham Farahani 2
1 Department of Computer Engineering, Faculty of Computer Science, Iranian eUniversity, Tehran, Iran.
2 Assistant Professor, Faculty of Computer Engineering, Iranian eUniversity, Tehran, Iran.
10.22034/jcse.2026.599524.1095
Abstract
Machine learning lifecycle security faces challenges beyond traditional software due to its dependence on data, models, software dependencies, and execution environments. This study proposes a dynamic framework for assessing machine learning-specific threats within enterprise MLSecOps. Security outputs from various tools across the ML lifecycle were collected, normalized, and transformed into unified risk indicators. Risk was assessed using five dimensions: impact severity, likelihood of occurrence, detection difficulty, organizational impact, and attacker accessibility. The resulting scores were dynamically updated according to changes in system components. To determine risk dimension values, a questionnaire was distributed to 17 experts, yielding 12 valid responses. The framework was evaluated using six scenarios: non-TLS communication, data version modification, insecure model files, dependency vulnerabilities, data drift, and prompt injection. Results identified prompt injection (92) and insecure model files (71) as the highest-risk scenarios, demonstrating the framework’s effectiveness in converting heterogeneous security evidence into dynamic risk assessments.
Keywords
MLSecOps
Dynamic Risk Assessment
Machine Learning Security
Risk Management
Security Threats
The CSI Journal on Computer Science and Engineering

Articles in Press, Accepted Manuscript
Available Online from 29 August 2026