Efficient Triple Modular Redundancy for Reliability Enhancement of DNNs Using Explainable AI

DNNs need protection in critical applications: Deep Neural Networks are increasingly used in safety-critical domains (like autonomous vehicles or medical diagnosis), where failures can have serious consequences. This paper addresses how to make them more reliable when hardware errors occur.

Explainable AI helps identify what matters most: Rather than protecting all parts of a neural network equally (which is expensive), the researchers use an Explainable AI technique called Layer-wise Relevance Propagation (LRP) to identify which individual weights and neurons are most important to the network's decisions. This allows them to focus protection efforts where it matters most.

Selective protection is more efficient: By using LRP to pinpoint the most critical components, the method achieves over 60% reliability improvement against hardware bit-flip errors in the AlexNet model while using the same computational overhead as existing approaches—demonstrating that smart targeting beats blanket protection.

Explainability serves a practical engineering purpose: This work shows that XAI methods aren't just for understanding AI decisions; they can be directly applied to solve real engineering problems like hardware fault tolerance, bridging the gap between interpretability research and practical reliability engineering.

Trade-offs between protection and efficiency matter: The paper demonstrates that responsible AI deployment in critical systems requires balancing reliability improvements against computational costs—a key consideration for real-world applications where resources are limited.