Deep neural networks (DNNs) are state-of-the-art algorithms for multiple\napplications, spanning from image classification to speech recognition. While\nproviding excellent accuracy, they often have enormous compute and memory\nrequirements. As a result of this, quantized neural networks (QNNs) are\nincreasingly being adopted and deployed especially on embedded devices, thanks\nto their high accuracy, but also since they have significantly lower compute\nand memory requirements compared to their floating point equivalents. QNN\ndeployment is also being evaluated for safety-critical applications, such as\nautomotive, avionics, medical or industrial. These systems require functional\nsafety, guaranteeing failure-free behaviour even in the presence of hardware\nfaults. In general fault tolerance can be achieved by adding redundancy to the\nsystem, which further exacerbates the overall computational demands and makes\nit difficult to meet the power and performance requirements. In order to\ndecrease the hardware cost for achieving functional safety, it is vital to\nexplore domain-specific solutions which can exploit the inherent features of\nDNNs. In this work we present a novel methodology called fault-aware training\n(FAT), which includes error modeling during neural network (NN) training, to\nmake QNNs resilient to specific fault models on the device. Our experiments\nshow that by injecting faults in the convolutional layers during training,\nhighly accurate convolutional neural networks (CNNs) can be trained which\nexhibits much better error tolerance compared to the original. Furthermore, we\nshow that redundant systems which are built from QNNs trained with FAT achieve\nhigher worse-case accuracy at lower hardware cost. This has been validated for\nnumerous classification tasks including CIFAR10, GTSRB, SVHN and ImageNet.\n