Towards Certification of Machine Learning-Based Distributed Systems

Marco Anisetti, Claudio A. Ardagna, Nicola Bena, Ernesto Damiani

Abstract

Traditional software certification methods, designed for deterministic systems, are inadequate for the non-deterministic nature of Machine Learning (ML). This paper introduces a novel multi-factor certification framework that evaluates the integrity of training data, the rigor of the development process, and the operational behavior of the final model. By shifting from a single-point binary check to a holistic audit of the ML lifecycle, this approach provides a standardized path toward certifying cloud-native, ML-driven applications.

Introduction: The Certification Crisis in ML

Historically, software certification relied on verifying that code adhered to deterministic "if-then" logic, as seen in legacy standards like the "Orange Book." However, the rise of ML has precipitated a "second crisis of certification." Modern applications increasingly rely on "black box" models whose behavior is shaped not just by code, but by the data they ingest and the specific parameters of their training.

Current certification schemes often treat ML models as static components, failing to account for vulnerabilities like training data poisoning or a lack of robustness against adversarial attacks. To address this, a framework is needed that can certify the non-deterministic components of distributed systems throughout their entire lifecycle.

The Multi-factor Certification Framework

The authors propose a Multi-factor Certification Scheme that decomposes the "Target of Certification" (ToC) into three distinct factors:

  1. Factor Data ($f_d$): Focuses on the dataset used for training and validation. It audits for data poisoning and ensures the data is balanced and representative of real-world scenarios.
  2. Factor Process ($f_p$): Evaluates the training methodology. This involves auditing training code and checkpoints to verify the use of defensive techniques such as randomized smoothing or adversarial training.
  3. Factor Model ($f_m$): Assesses the model's behavior during inference. This traditional step involves testing the model against adversarial samples to verify properties like robustness and privacy.
The proposed certification workflow.

This approach moves certification from a simple pass/fail check of a binary file to a comprehensive audit. It acknowledges that in ML, properties like "robustness" are often defined by the specific verification methods used against them.

Case Study: Malware Detection

The framework was evaluated using a malware detector as a reference scenario. The training data consisted of real-world performance metrics and synthetic data generated via Generative Adversarial Networks (GANs). The model was then tested against subtly perturbed malware activities designed to evade detection.

Key Results

The application of the three-factor scheme revealed critical insights into the detector's failures. While the data factor passed, both the process and model factors failed, highlighting that a clean dataset is insufficient if the training process is flawed.

Evaluation results of the three-factor certification scheme.

As shown in the table, the model factor ($f_m$) failed because the recall dropped to zero under adversarial attack. Crucially, the process factor ($f_p$) audit identified the root cause: the developers had omitted randomized smoothing and adversarial training. This transparency allows regulators to trace failures back to specific stages of the ML lifecycle.

Discussion and Future Directions

The "three-factor" approach provides a necessary "pedigree" for ML models, ensuring they are not merely "houses of cards" that perform well only in controlled environments. However, several challenges remain:

  • Recursive Complexity: ML applications often involve complex pipelines where models are chained together. Developing formal methods to compose certificates for these multi-stage systems is a priority.
  • Continuous Certification: Because ML models drift over time as they encounter new data, certification must evolve from a one-time event into a continuous monitoring process.
  • Advanced Verification: Future iterations of the framework should integrate mathematical proofs of behavior (abstract interpretation) and surrogate white-box models to improve the explainability of black-box decisions.

By providing a structured way to audit data, process, and model performance, this framework represents a significant step toward trustworthy and certifiable ML-based distributed systems.

Summary updated

BibTeX
@article{anisetti2023towards,
  author = {Marco Anisetti and Claudio A. Ardagna and Nicola Bena and Ernesto Damiani},
  title = {{Towards Certification of Machine Learning-Based Distributed Systems}},
  journal = {IEEE Internet Comput},
  year = {2023},
  doi = {10.1109/mic.2023.3322327},
}
Figure