An Empirical Investigation Into Benchmarking Model Multiplicity for Trustworthy Machine Learning: A Case Study on Image Classification

Prakhar GaneshWACV 2024

TL;DR

This paper introduces "multiplicity sheets," a standardized framework to benchmark model multiplicity across various trustworthy machine learning metrics. They demonstrate its utility with an image classification case study, revealing that even with explicit specifications, model multiplicity persists, posing a significant challenge for reliable AI deployment.

What Problem Does This Paper Address?

Deep learning models are incredibly powerful, but their success often comes from being "over-parameterized." This leads to a phenomenon called "model multiplicity," where many different models can achieve similar performance on a task but behave very differently under the hood. This hidden variability is a major roadblock for "trustworthy machine learning" – ensuring AI systems are fair, robust, private, and secure. Prior research often looked at these trustworthiness aspects in isolation, making it hard to get a holistic view of multiplicity. This paper aims to bridge that gap by providing a unified way to measure and compare model multiplicity across different trustworthiness dimensions.

Methods

The core contribution is a novel framework for benchmarking model multiplicity.

  • Accuracy Under Intervention: To compare diverse trustworthiness metrics (like fairness, robustness, privacy, security) on a common scale, the authors propose translating each metric into an "accuracy under intervention." This means designing specific tests or perturbations that allow us to measure model accuracy in a way that proxies the original trustworthiness metric. For example, instead of measuring the minimum adversarial distance to flip a label (security), they measure accuracy under a fixed adversarial distance budget.
  • Multiplicity Sheets: These are a structured way to record and visualize the "accuracy under intervention" across different model configurations (e.g., varying hyperparameters, random seeds, architectures). They offer three levels of readability: raw scores, aggregated differences (Δmax) to highlight trends, and an overall multiplicity score (Δall max) for a single-value comparison.

Datasets

  • UTKFace: Used as the primary dataset for the image classification case study, containing facial images labeled with gender, race, and age.
  • FairFace: Used for evaluating out-of-distribution (OOD) robustness.
  • CelebA: Used for evaluating robustness under "unforeseen circumstances."

Key Insights and Ideas

  • A Unified Language for Multiplicity: The "accuracy under intervention" and "multiplicity sheets" provide a much-needed common ground for discussing and comparing model multiplicity across different trustworthiness aspects.
  • Multiplicity is Pervasive: The case study demonstrates that multiplicity isn't just an issue for accuracy; it significantly impacts fairness, robustness, privacy, and security.
  • Random Seeds Matter (Especially for Fairness): While hyperparameters and architecture choices influence multiplicity, random seeds can have a substantial impact, particularly on fairness metrics.
  • Intersectional Groups Amplify Fairness Issues: The severity of fairness multiplicity is even more pronounced when considering combinations of protected attributes (e.g., age and race).
  • Trustworthiness Metrics Have Higher Multiplicity: The paper shows that trustworthiness metrics (fairness, robustness, privacy, security) exhibit significantly higher multiplicity ranges (Δall max) compared to standard accuracy.
  • Specifications Aren't a Silver Bullet: Even after applying rigorous model selection based on multiple trustworthy metrics, unforeseen circumstances can still reveal significant multiplicity, highlighting the deep-rooted nature of the problem.

Key Results

  • Accuracy Multiplicity (Δall max): 1.12 on UTKFace.
  • Group Fairness Multiplicity (Asian, Δall max): 3.24 on UTKFace, nearly three times higher than accuracy multiplicity.
  • OOD Robustness Multiplicity (FairFace, Δall max): 3.51 on UTKFace, comparable to fairness multiplicity.
  • Privacy Multiplicity (Perturbation Accuracy, Δall max): 5.30 on UTKFace, almost five times higher than accuracy multiplicity.
  • Security Multiplicity (PGD Accuracy, Δall max): 5.53 on UTKFace, also significantly higher than accuracy multiplicity.
  • The paper demonstrates that even after selecting models that meet high standards for fairness, robustness, privacy, and security, significant multiplicity can still exist for "unforeseen" metrics.

Limitations and Future Work

  • Scope of Interventions: While the paper proposes "accuracy under intervention," the specific interventions are tailored to classification tasks. Extending this to other domains like regression or generation would require further work.
  • Standardization of Multiplicity Sheets: The authors acknowledge that further research is needed to develop even more effective methods for recording and standardizing multiplicity sheets.
  • Beyond Checklists: The paper strongly advocates for moving beyond simply adding trustworthy metrics as a "checklist" during model selection. Future work should focus on more fundamental investigations into model multiplicity to address potential unforeseen failures in real-world deployments and build truly trustworthy systems.

Summary updated

BibTeX
@inproceedings{ganesh2024empirical,
  author = {Prakhar Ganesh},
  title = {{An Empirical Investigation Into Benchmarking Model Multiplicity for Trustworthy Machine Learning: A Case Study on Image Classification}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2024},
}
Figure