אסף שבתאי

The widespread adoption of machine learning (ML) systems has intensified concerns about their security and the rise of adversarial machine learning (AML) techniques that exploit inherent model vulnerabilities. This highlights the need for comprehensive risk-assessment methods specifically tailored to ML-based systems. Traditional cybersecurity frameworks are not equipped to address the distinctive challenges of adversarial threats, while existing AML evaluations focus mainly on technical robustness and often overlook real-world factors such as deployment environments, system dependencies, and attack feasibility. Previous efforts toward comprehensive AML risk assessment have remained domain-specific, limiting their applicability across diverse systems. To address these gaps, we present FRAME, a comprehensive framework for assessing AML risks in ML systems. FRAME quantifies adversarial risk across three dimensions: the target system's deployment environment, the characteristics of AML techniques, and empirical findings from prior research. It integrates a structured dataset of AML attacks, enabling context-aware and reproducible evaluation, and includes a semi-automated scoring process with LLM-assisted customization for system-specific assessments. Validated across six real-world use cases, including e-commerce feedback scoring and email malware detection, FRAME achieved an average overall accuracy of 9/10, with expert assessments indicating high attack-specific accuracy (9.2/10) and relevance (8.9/10), demonstrating strong alignment with expert analyses. By enabling organizations to identify and prioritize adversarial risks, FRAME provides a practical and scalable foundation for secure and informed deployment of ML systems.