[Paper Review] DeepGD: A Multi-Objective Black-Box Test Selection Approach for Deep Neural Networks

Zohreh Aghababaeyan, Manel Abdellatif+2 more|arXiv (Cornell University)|Mar 8, 2023

Adversarial Robustness in Machine Learning70 references4 citations

TL;DR

DeepGD is a black-box, multi-objective test selection approach for deep neural networks that prioritizes unlabeled inputs with high uncertainty and diversity to maximize fault detection while minimizing labeling costs. It outperforms existing white-box and black-box methods in fault revelation and improves model retraining accuracy by selecting informative test inputs.

ABSTRACT

Deep neural networks (DNNs) are widely used in various application domains such as image processing, speech recognition, and natural language processing. However, testing DNN models may be challenging due to the complexity and size of their input domain. Particularly, testing DNN models often requires generating or exploring large unlabeled datasets. In practice, DNN test oracles, which identify the correct outputs for inputs, often require expensive manual effort to label test data, possibly involving multiple experts to ensure labeling correctness. In this paper, we propose DeepGD, a black-box multi-objective test selection approach for DNN models. It reduces the cost of labeling by prioritizing the selection of test inputs with high fault revealing power from large unlabeled datasets. DeepGD not only selects test inputs with high uncertainty scores to trigger as many mispredicted inputs as possible but also maximizes the probability of revealing distinct faults in the DNN model by selecting diverse mispredicted inputs. The experimental results conducted on four widely used datasets and five DNN models show that in terms of fault-revealing ability: (1) White-box, coverage-based approaches fare poorly, (2) DeepGD outperforms existing black-box test selection approaches in terms of fault detection, and (3) DeepGD also leads to better guidance for DNN model retraining when using selected inputs to augment the training set.

Motivation & Objective

• To address the high cost of labeling large unlabeled datasets in DNN testing, especially when model internals are inaccessible.
• To reduce reliance on white-box coverage-based test selection, which often fails to correlate with actual mispredictions.
• To develop a black-box test selection method that leverages uncertainty and diversity to maximize fault-revealing power.
• To improve model retraining by selecting test inputs that enhance generalization and fault detection.
• To provide a cost-effective, one-time labeling solution that avoids iterative human-in-the-loop labeling required by active learning.

Proposed method

• DeepGD uses model output uncertainty scores to identify inputs likely to be mispredicted, based on low confidence or high entropy in predictions.
• It computes diversity by measuring the distance between input embeddings in the last hidden layer, ensuring selected inputs cover different regions of the input space.
• A multi-objective optimization framework balances high uncertainty and high diversity to select test inputs that maximize fault detection.
• The method operates without access to model internals or training data, making it applicable to third-party or proprietary DNNs.
• It employs a search-based approach to explore the unlabeled dataset and select a compact, high-impact test suite under a fixed budget.
• The selected inputs are used for both testing and retraining, improving model robustness with minimal labeling effort.

Experimental results

Research questions

• RQ1 Can a black-box test selection approach that combines uncertainty and diversity outperform white-box coverage-based methods in fault detection for DNNs?
• RQ2 How does the fault-revealing ability of DeepGD compare to existing black-box test selection techniques under the same labeling budget?
• RQ3 To what extent does using DeepGD-selected inputs for retraining improve DNN model accuracy and fault tolerance?
• RQ4 Does the combination of uncertainty and diversity lead to better fault detection than using either criterion alone?
• RQ5 Is DeepGD effective across diverse DNN architectures and datasets, including those with complex input distributions?

Key findings

• DeepGD significantly outperforms white-box, coverage-based test selection approaches in fault detection, despite their reliance on internal model metrics.
• DeepGD achieves higher fault-revealing power than existing black-box methods by effectively balancing uncertainty and diversity in test input selection.
• The approach leads to better model retraining outcomes, with improved accuracy when using selected test inputs to augment the training set.
• The second-best test selection method varies across models and datasets, indicating that no single approach universally dominates, which supports the robustness of DeepGD's multi-objective strategy.
• DeepGD's performance is consistent across four widely used datasets and five DNN models, demonstrating its generalizability.
• The method reduces labeling costs by focusing on a small, high-impact subset of unlabeled data, making it practical for real-world testing with limited budgets.