Undetected out-of-distribution (OOD) inputs pose a significant threat to the reliability of deep learning models, as they may lead to unexpected behaviors during inference. Several studies have proposed effective OOD input detection methods. However, soft errors—another significant threat to reliability—can impact both the classification results of neural network models and the ID/OOD detections of OOD detection methods. To provide a resilient OOD detection solution against soft errors, we analyze the effect of soft errors on neural network models with gradient-based input perturbation (GIP) approaches, which are representative methods for OOD detection. Building on our analysis, we propose ProGIP, which incorporates two software-level range-based fault detectors to protect all execution phases of GIP approaches, including two forward passes and one backward pass. Because it is purely software‑based and adds just two scalar comparisons, ProGIP is readily deployable even on resource‑constrained embedded platforms. Our ProGIP solution enables GIP approaches to distinguish between ID, OOD, and fault-affected inferences, detecting 97.7% of critical faults with a negligible runtime overhead of only 0.84%. Experimental results with 2.4 million fault injections across various neural networks and OOD detection methods demonstrate ProGIP’s effectiveness in ensuring comprehensive reliability against non-malicious threats.

As the deployment of neural networks in safety-critical applications proliferates, it becomes imperative that they exhibit consistent and dependable performance amidst hardware malfunctions. Several protection schemes have been proposed to protect neural networks, but they suffer from huge overheads or insufficient fault coverage. This paper presents Maintaining Sanity, a comprehensive and efficient protection technique for CNNs. Maintaining Sanity extends the state-of-the-art algorithm-based fault tolerance for CNN, utilizing hamming codes and checkpointing to correct over 99.6% of critical faults with about 72% runtime overhead and minimal memory overhead compared to traditional triple modular redundancy (TMR) techniques.