A Multi-Layer Behavioral and Explainable Framework for Robust Detection of Backdoor Attacks in Deep Neural Networks
Backdoor attacks pose a critical threat to Deep Neural Networks (DNNs) by embedding hidden behaviors that are activated only under specific trigger conditions, compromising the reliability of Artificial Intelligence (AI) systems. Existing detection approaches often rely on single-method assumptions, limited data access, or controlled environments, limiting their effectiveness against adaptive, real-world attacks. To address these limitations, this study proposes a multi-layer, explainable framework for robust backdoor detection in DNNs. The approach integrates complementary detection mechanisms — activation clustering, spectral-signature analysis, Gaussian Mixture Models, entropy-based evaluation, and input perturbation — within a unified pipeline. Each layer captures distinct indicators of anomalous behavior, enabling comprehensive analysis across structural, statistical, and behavioral dimensions. An explainable component provides interpretable insights into detection decisions. Experimental evaluation on the MNIST and CIFAR-10 datasets demonstrates that the framework achieves 97–99% detection rates with false positive rates (FPRs) below 2%, while reducing attack success rates (ASRs) by over 94% across diverse trigger types. The results confirm that combining multiple detection perspectives significantly improves robustness compared with single-layer defenses. Overall, this work advances AI security by introducing a scalable, practical defense mechanism that operates under limited-knowledge conditions and supports trustworthy deployment in real-world environments.