Framing Adversarial Machine Learning and Federated Learning Threats through MITRE ATLAS

Abstract

As the adoption of Federated Learning (FL) accelerates across sectors prioritizing privacy, its decentralized architecture introduces novel cybersecurity threats that remain underrepresented in existing adversarial threat taxonomies. This paper bridges this gap by systematically analyzing FL-specific adversarial techniques and mapping them to the MITRE ATLAS (Adversarial Threat Landscape for Artificial-Intelligence Systems) framework, a living knowledge base for Artificial Intelligence Systems threats. Through a structured methodology and systematic literature review of 126 peer-reviewed articles published between 2021-2025, complemented by empirical validation through detailed case studies, we were able to find out that federated learning is vulnerable to several critical vulnerabilities, such as model poisoning, privacy leakage, and collusion attacks both in cross-silo and cross-device settings. The in-depth analysis of the current coverage in MITRE ATLAS reveals considerable weaknesses in its coverage and the mitigation measures are critically analyzed in the light of computational overhead, scalability concerns, and regulatory compliance issues. This contribution proposes extensions to the MITRE ATLAS framework, enables AI threat intelligence operationalization and provides a systematized roadmap of standardization of federated learning threat modeling in the ATLAS framework.