Modeling Threats to AI-ML Systems Using STRIDE

Lara Mauri, Ernesto DamianiSensors 2022 · DOI

TL;DR

This paper introduces STRIDE-AI, a novel threat modeling methodology specifically designed for AI-ML systems. By extending the well-established STRIDE framework, STRIDE-AI provides a structured way for security practitioners to identify and analyze vulnerabilities across the entire ML lifecycle, from data management to model deployment and maintenance. This is crucial for building more secure and trustworthy AI-powered applications.

What Problem Does This Paper Address?

The increasing integration of Artificial Intelligence (AI) and Machine Learning (ML) into critical infrastructures brings significant security risks. Unlike traditional IT systems, ML models have unique vulnerabilities, such as susceptibility to data poisoning or adversarial attacks during training and inference. Existing threat modeling approaches often fall short in addressing the complexities of the ML lifecycle, leaving security practitioners without a clear, systematic way to identify and mitigate these novel threats. This gap hinders the development of secure AI-ML systems.

Methods

The core contribution of this paper is the development of STRIDE-AI, an adaptation of Microsoft's STRIDE (Spoofing, Tampering, Repudiation, Information Disclosure, Denial of Service, Elevation of Privilege) threat modeling framework for AI-ML systems. The authors propose an asset-centered approach, first decomposing the ML system into key asset categories (Data, Models, Actors, Processes, Tools, Artefacts) and then applying Failure Mode and Effects Analysis (FMEA) to identify potential failure modes for each asset. These failure modes are then mapped to the STRIDE threats and the corresponding security properties (Authenticity, Integrity, Non-repudiation, Confidentiality, Availability, Authorization - CIA³-R) they endanger. This mapping allows for a systematic identification of threats specific to AI-ML components.

Datasets

The paper does not introduce or utilize specific datasets for training or evaluation in the context of demonstrating the STRIDE-AI methodology. Instead, it focuses on the conceptual framework and its application to a real-world use case.

Key Insights and Ideas

  • ML Systems Have Unique Attack Surfaces: The paper highlights that ML systems introduce attack vectors not present in traditional software, particularly concerning the manipulation of training data and model inference.
  • Lifecycle-Centric Threat Modeling is Essential: Threats can emerge at any stage of the ML lifecycle, from data collection and preprocessing to model training, tuning, deployment, and maintenance. A holistic approach is therefore necessary.
  • STRIDE Adaptation is Feasible and Beneficial: The STRIDE framework, when adapted to the AI-ML domain (STRIDE-AI), provides a robust and systematic way to categorize and understand threats.
  • Failure Modes Drive Threat Identification: By systematically analyzing how ML assets can fail (using FMEA), researchers and practitioners can effectively identify the specific threats that exploit these failures.
  • Security Properties are Key: Mapping failure modes to violations of CIA³-R security properties provides a clear link between asset vulnerabilities and the STRIDE threats that exploit them.

Key Results

The paper presents a detailed methodology (STRIDE-AI) for modeling threats to AI-ML systems. It provides:

  • A classification of ML assets into six macro-categories.
  • A guide for applying FMEA to AI-ML systems to identify failure modes.
  • A mapping of ML asset failure modes to CIA³-R security properties and corresponding STRIDE threats.
  • A real-world use case demonstrating the application of STRIDE-AI to a predictive maintenance system in the energy sector, identifying specific threats like Spoofing and Tampering for data streams and trained models.
  • A DREAD scorecard analysis for the identified threats in the use case, categorizing them as medium risk.

Limitations and Future Work

The authors acknowledge that:

  • No Universal Security Control Framework: While STRIDE-AI helps identify threats, there is a lack of a comprehensive security control framework specifically designed for ML models. Future work should focus on developing such controls and mapping them to ML functionalities and weaknesses.
  • Subjectivity in Threat Prioritization: While DREAD was used for preliminary threat assessment, its inherent subjectivity can lead to inconsistent results. The authors suggest exploring alternative or complementary prioritization methods.
  • Continuous Updating is Necessary: The dynamic nature of ML systems and evolving threats means that threat modeling must be an ongoing process, requiring continuous updates based on new information and research.

Future work could involve further refining STRIDE-AI, developing ML-specific security controls, and exploring how technologies like Distributed Ledger Technologies (DLTs) can enhance the security of ML systems.

Summary updated

BibTeX
@article{mauri2022modeling,
  author = {Lara Mauri and Ernesto Damiani},
  title = {{Modeling Threats to AI-ML Systems Using STRIDE}},
  journal = {Sensors},
  year = {2022},
  doi = {10.3390/s22176662},
}
Figure