Work

CD-BD: Interpretable Backdoor Model Detection via Concept-Level Dissimilarities

Year: 2026
Type: article
Abstract: Image-classification models have been adopted as visual components in AI-assisted education applications, such as educational visual content analysis and interactive learning robots. However, their security can be compromised by backdoor attacks, in which a compromised model behaves normally on clean inputs but produces attacker-targeted predictions when activated by hidden triggers. This poses a serious threat to the reliability of educational applications and highlights the need for effective backdoor detection approaches before model deployment. Despite achieving promising results, existing backdoored model detection methods provide limited insight into how backdoor injection changes the model’s concept-level decision basis, making the detection process insufficiently interpretable. To address this problem, this paper proposes CD-BD, a general interpretable backdoored model detection method based on concept-level dissimilarities. Through ACE- and TCAV-based analysis, we observe that clean and backdoored models exhibit different concept-interpretation behavior: backdoored models tend to identify concepts that are more closely related to trigger-like patterns, whereas clean models rely on more natural class-related concepts. Based on this phenomenon, CD-BD filters reliable concepts by TCAV score and statistical significance, constructs concept-count features across classes and repeated runs, extracts color, texture, and edge descriptors from representative concept images, and concatenates these interpretable cues into composite model-level feature vectors for SVM classification. Experiments on CIFAR-10 with PreActResNet-18 and BadNets show that CD-BD achieves 96% detection accuracy, 100% TPR, and 7% FPR using five-fold cross-validation, compared with 93% accuracy, 80% TPR, and 0% FPR for BAN. These results offer preliminary evidence that concept-level dissimilarities provide an effective and human-understandable basis for detecting backdoored image-classification models.
Language: English
Open Access status: gold
Estimated APC paid: $2,896
OpenAlex ID: W7226037889
FWCI:
Cites: 31
Cited by:
Related to: 10