CAFD: Concept-Aware DNN Fault Detection using VLMs

Background

Fault detection for Deep Neural Networks (DNNs) is critical for reliable AI systems. Existing hybrid approaches that combine multiple information sources achieve better detection but are often hampered by high computational overhead, limiting their practicality for real-world, large-scale deployment. This creates a need for methods that are both accurate and efficient.

Key Points

• Core Innovation: Concept-Aware Fault Detection (CAFD). This is a learning-based framework designed to integrate multiple information sources effectively without incurring prohibitive computational costs.
• Novel Feature: Concept Failure Ratio (CFR). This is the central contribution. CFR moves beyond raw numerical outputs to leverage semantic information.
  • It uses Vision-Language Models (VLMs) to extract textual concepts from input images.
  • It then quantifies the likelihood that the presence of these concepts is correlated with the DNN's incorrect predictions (failures).
• Integrated Feature Set: CAFD trains on a carefully selected combination of:
  1. Model-based signals: Derived from the DNN's own output probabilities.
  2. Distance-based features: Likely measuring similarity in the model's feature space.
  3. The concept-based CFR feature: Providing the new semantic layer.
• Empirical Superiority: CAFD was rigorously tested against five state-of-the-art baselines on three different DNN models and datasets, including the large-scale ImageNet. Across all tests and under constrained selection budgets (a practical scenario where only a limited number of predictions can be verified), CAFD consistently achieved the highest Fault Detection Rate (FDR), with an average improvement of 18.3%.

Significance

• Addresses a Key Trade-off: CAFD directly tackles the core tension between detection performance and computational efficiency in DNN fault detection. It achieves superior accuracy while maintaining practical scalability.
• Introduces Semantic Reasoning: The CFR metric represents a significant shift. By incorporating high-level semantic concepts via VLMs, it provides a complementary dimension to fault detection that purely numerical methods might miss, leading to more effective identification of failure-prone scenarios.
• Practical Impact: The consistent performance gains across diverse models and datasets, especially under constrained budgets, demonstrate CAFD's potential as a deployable tool for improving the safety and reliability of DNN applications in real-world settings.