Core Concepts
DeepKnowledge enhances DNN robustness by identifying transfer knowledge neurons for improved generalization and testing adequacy.
Abstract
The article introduces DeepKnowledge, a systematic testing methodology for DNN-based systems. It addresses the lack of systematic testing approaches that assess DNN's capability to generalize beyond training data distribution. DeepKnowledge focuses on Transfer Knowledge neurons to enhance DNN robustness and reduce residual risk. The methodology provides objective confidence measurement on testing activities given data distribution shifts. Empirical evaluations demonstrate its effectiveness in detecting adversarial attacks and improving coverage criteria.
-
Introduction
- Despite DNN success, they are fragile to data distribution shifts.
- Safety incidents highlight the need for dependable DNNs.
-
Background
- Overview of DNN-based systems and software testing methodologies.
-
DeepKnowledge
- Identifies Transfer Knowledge neurons for generalization.
- Utilizes Hellinger Distance to measure knowledge change.
-
Implementation and Hardware
- Experiments conducted on high-performance computers with GPU clusters.
-
Experimental Evaluation
- Research questions focus on knowledge generalization, hyperparameter sensitivity, effectiveness, correlation with other criteria, and OOD dataset impact.
-
Results and Discussion
- DeepKnowledge improves DNN accuracy through guided retraining.
- Hyperparameter sensitivity affects coverage values and test set diversity.
Stats
"We report improvements of up to 10 percentage points over state-of-the-art coverage criteria for detecting adversarial attacks on several benchmarks, including MNIST, SVHN, and CIFAR."
Quotes
"DeepKnowledge posits that core computational DNN units can generalize under domain shift."
"Our empirical evaluation demonstrates the usefulness and effectiveness of DeepKnowledge."