CAFD: Concept-Aware DNN Fault Detection using VLMs

Fault detection for Deep Neural Networks (DNNs) has received increasing attention in recent years. While more advanced hybrid approaches have been proposed to combine multiple sources of information and outperform earlier techniques, they often incur substantial computational overhead, limiting scalability and practicality in real-world settings. In this paper, we introduce Concept-Aware Fault Detection (CAFD), a learning-based approach that achieves superior fault detection performance by effectively integrating multiple information sources while maintaining practical efficiency. Specifically, CAFD is trained using a carefully selected set of informative features, including model-based signals derived from the DNN's outputs, distance-based features, and a novel concept-based feature, called Concept Failure Ratio (CFR). CFR leverages Vision-Language Models (VLMs) to extract textual concepts from images and quantify the likelihood that their presence is associated with DNN failures. By incorporating this feature, CAFD benefits from complementary semantic information, enabling more effective fault detection. Our results demonstrate that CFR serves as an effective indicator for DNN fault detection. We conduct an extensive empirical evaluation of CAFD, comparing it against five state-of-the-art baselines across three subject DNN models and datasets, including ImageNet. Across a wide range of constrained selection budgets, CAFD consistently outperforms all baselines in Fault Detection Rate (FDR), achieving average FDR improvements of 18.3% across all investigated subjects and budget sizes.

📋 한 줄 요약

**[DNN Fault Detection / VLM]** CAFD가 model-based signal·distance feature·VLM 기반 Concept Failure Ratio(CFR)를 통합 — ImageNet 등 3 DNN·5 SOTA baseline 대비 평균 FDR 18.3% 향상, hybrid 방식 대비 계산 부담 완화.

🎯 핵심 기여도

DNN의 fault detection이 최근 주목받으며 multiple information source를 결합한 advanced hybrid 접근이 earlier 기법을 outperform하지만, substantial 계산 overhead로 scalability·practicality가 제약됨을 진단.
CAFD(Concept-Aware Fault Detection) 도입 — learning 기반 접근으로 multiple information source를 효과적으로 통합하면서 practical 효율 유지.
학습 입력으로 carefully selected informative feature 사용 — DNN output 유도 model-based signal·distance-based feature·novel concept-based feature(Concept Failure Ratio, CFR).
CFR은 VLM을 활용해 이미지에서 textual concept을 추출하고, 그 presence가 DNN failure와 연관될 likelihood를 정량화 — complementary semantic information을 더 효과적 fault detection으로 연결.

💡 핵심 아이디어

DNN fault detection 성능은 model-based·distance-based feature를 넘어 image-level semantic concept(VLM이 추출하는 textual concept)이 failure와 어떻게 연관되는지를 정량화하는 CFR과 결합할 때 효율·정확도를 동시에 개선할 수 있다.

🔬 기술적 접근법

**방법론**: CAFD — Concept-Aware Fault Detection (learning 기반 통합).
**핵심 기법**: (1) Model-based signal (DNN output 유도 feature), (2) Distance-based feature, (3) Novel Concept Failure Ratio(CFR) — VLM이 추출한 textual concept과 DNN failure의 연관 likelihood 정량화, (4) 세 source를 carefully selected feature set으로 결합해 학습, (5) 광범위 budget 범위에서 5 SOTA baseline 비교.

📊 주요 결과

CFR이 DNN fault detection의 효과적 indicator임을 입증.
3 DNN·데이터셋(ImageNet 포함)·5 SOTA baseline 대비 광범위 평가.
다양 constrained selection budget에서 CAFD가 모든 baseline outperform.
모든 subject·budget 평균 FDR 18.3% 개선.

💭 의의 및 한계

**의의**: VLM 의미 정보를 fault detection에 효과적 결합, hybrid 방식의 계산 부담 완화하면서 성능 향상, ImageNet 규모 평가로 실용성 입증. **한계**: VLM concept 추출 비용·품질에 성능 의존, CFR 정의에 사용된 concept vocabulary 의존, 매우 큰 DNN에서의 scaling은 추가 검증.

🚀 실용적 활용

안전 critical 시스템의 DNN 신뢰성 검증.
실제 배포 시 분류 실패 사전 탐지.
ML pipeline의 quality gate·monitoring.