DOC: Detection of On-Line Failures in CNNs

The integration of Artificial Intelligence (AI) in safety-critical systems raises concerns about reliability, particularly due to the inherent uncertainty of AI algorithms and the complexity of modern hardware, compromising billion of transistors. Existing solutions, such as Algorithm-Based Fault Tolerance often focus on running detection algorithms after every inference, introducing a not negligible overhead in the detection phase. This paper introduces a two-phase fault detection technique for Convolutional Neural Networks (CNNs) with floating-point precision. The first phase identifies easily detectable faults (such as those stemming from a bit-flip on the 30th of a floating-point representation), while the second one targets hard-to-detect critical faults—those producing a wrong prediction but having no visible effect during faults' propagation. Although these faults constitute only 1.4% of all critical faults, their detection is crucial for ensuring system reliability. Validated on the CIFAR-10 dataset with a ResNet-20 model, the proposed method achieves up to 99.67% coverage of critical inferences while maintaining moderate computational overhead. This lightweight, real-time solution enhances the robustness of CNNs in safety-critical applications.