Do-AIQ: A Design-of-Experiment Approach to Quality Evaluation of AI Mislabel Detection Algorithm

J. Lian, K. Choi, Balaji Veeramani, A. Hu, L. Freeman, E. Bowen, X. DengarXiv 2022 · OpenReview

Abstract

The quality of Artificial Intelligence (AI) algorithms is of significant importance for confidently adopting algorithms in various applications such as cybersecurity, healthcare, and autonomous driving. This work presents a principled framework of using a design-of-experimental approach to systematically evaluate the quality of AI algorithms, named as Do-AIQ. Specifically, we focus on investigating the quality of the AI mislabel data algorithm against data poisoning. The performance of AI algorithms is affected by hyperparameters in the algorithm and data quality, particularly, data mislabeling, class imbalance, and data types. To evaluate the quality of the AI algorithms and obtain a trustworthy assessment on the quality of the algorithms, we establish a design-of-experiment framework to construct an efficient space-filling design in a high-dimensional constraint space and develop an effective surrogate model using additive Gaussian process to enable the emulation of the quality of AI algorithms. Both theoretical and numerical studies are conducted to justify the merits of the proposed framework. The proposed framework can set an exemplar for AI algorithm to enhance the AI assurance of robustness, reproducibility, and transparency.

1. Introduction

As AI moves into high-stakes domains like cybersecurity and healthcare, "AI assurance"—the guarantee of robustness and reproducibility—is becoming a critical bottleneck. Specifically, Mislabel Detection (MLD) algorithms are designed to identify "poisoned" or incorrectly labeled data. However, these algorithms themselves are sensitive to their own hyperparameters and the quality of the data they process, such as class imbalance.

Evaluating these systems is difficult because the "search space" of possible conditions is massive and often constrained (e.g., class proportions must sum to 100%). Randomly testing a few configurations is insufficient for understanding an algorithm's limitations. The Do-AIQ framework addresses the lack of a systematic, statistically rigorous framework for evaluating AI quality under complex, intertwined conditions.

2. The Do-AIQ Framework

The Do-AIQ framework treats the AI evaluation process like a physical laboratory experiment, systematically mapping the relationship between input factors and performance metrics.

The Do-AIQ framework for systematic AI algorithm evaluation using Design-of-Experiment (DoE) and emulators.

Key components of the framework include:

  • Space-Filling Design: To explore the high-dimensional space of class proportions and hyperparameters, the framework uses a "modified maximin criterion." This ensures that test points are spread out as much as possible within the linear constraints of a simplex.
  • Coordinate Exchange Optimization: An efficient algorithm combines discrete coordinate exchange and continuous optimization to find representative design points without exhausting computational resources.
  • Additive Gaussian Process (AGP) Surrogate: Since training an AI model thousands of times is computationally expensive, the framework uses an AGP as a "surrogate model" or emulator. This statistical model predicts AI performance at any given point in the parameter space, providing both a prediction and a measure of uncertainty.
  • Hybrid Factor Handling: The framework uniquely manages a mix of continuous factors (e.g., loss function weights), categorical factors (e.g., different datasets), and constrained factors (e.g., class proportions).

3. Experimental Methodology

The framework was validated using two computer vision benchmarks: MNIST (handwritten digits) and FashionMNIST (clothing items). These datasets were used to test the MLD algorithm's ability to identify injected mislabels across varying levels of class imbalance and poisoning.

Architecture of the Variational Autoencoder (VAE) structure utilized within the Mislabel Detection (MLD) algorithm.

The MLD algorithm utilized in this study relies on a Variational Autoencoder (VAE) architecture, as shown above, to distinguish between clean and noisy data by analyzing reconstruction errors and latent space distributions.

4. Results and Discussion

4.1 Surrogate Accuracy and Design Efficiency

The proposed space-filling design consistently outperformed the standard Kennard-Stone (K-S) algorithm in coverage and maximin metrics. Furthermore, the Additive Gaussian Process proved to be a superior emulator compared to traditional linear regression. For predicting detection accuracy, the AGP achieved a Mean Squared Error (MSE) of $6.11 \times 10^{-5}$, an order of magnitude improvement over the linear model's $0.0006$.

4.2 Key Insights into AI Robustness

  • The "Conservative" Threshold: When MLD detection accuracy is low, it is also highly unstable. However, once performance crosses a threshold of approximately 0.7 accuracy, the algorithm becomes significantly more stable and predictable.
  • MLD vs. CNN Sensitivity: While standard CNN classification accuracy is highly dependent on the specific dataset, the detection accuracy of the MLD algorithm was remarkably robust across different data types. This suggests MLD relies more on the statistical difference between clean and noisy data than on primary image features.
  • Linear Poisoning Effect: The proportion of mislabeled data affects detection accuracy in a nearly linear fashion. As poisoning increases, the ability to distinguish signal from noise degrades predictably.
  • Class Imbalance: While class imbalance negatively affects performance, it was found to be a secondary factor compared to the raw percentage of mislabeled observations.

5. Conclusion

Do-AIQ provides a principled approach to AI assurance. Future work will focus on scaling the design construction for datasets with hundreds of classes (e.g., CIFAR-100) and implementing sequential design to further reduce the number of required training runs. Additionally, extending the framework to evaluate fairness and bias in AI remains a promising direction for ensuring equitable performance across diverse subgroups.

Summary updated

Abstract

The quality of Artificial Intelligence (AI) algorithms is of significant importance for confidently adopting algorithms in various applications such as cybersecurity, healthcare, and autonomous driving. This work presents a principled framework of using a design-of-experimental approach to systematically evaluate the quality of AI algorithms, named as Do-AIQ. Specifically, we focus on investigating the quality of the AI mislabel data algorithm against data poisoning. The performance of AI algorithms is affected by hyperparameters in the algorithm and data quality, particularly, data mislabeling, class imbalance, and data types. To evaluate the quality of the AI algorithms and obtain a trustworthy assessment on the quality of the algorithms, we establish a design-of-experiment framework to construct an efficient space-filling design in a high-dimensional constraint space and develop an effective surrogate model using additive Gaussian process to enable the emulation of the quality of AI algorithms. Both theoretical and numerical studies are conducted to justify the merits of the proposed framework. The proposed framework can set an exemplar for AI algorithm to enhance the AI assurance of robustness, reproducibility, and transparency.
BibTeX
@misc{lian2022do,
  author = {J. Lian and K. Choi and Balaji Veeramani and A. Hu and L. Freeman and E. Bowen and X. Deng},
  title = {{Do-AIQ: A Design-of-Experiment Approach to Quality Evaluation of AI Mislabel Detection Algorithm}},
  year = {2022},
  eprint = {2208.09953},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2208.09953},
}
Figure