A Low-cost Fault Corrector for Deep Neural Networks through Range Restriction

With the increasing adoption of deep neural networks (DNNs) in safety-critical domains such as autonomous vehicles and industrial robotics, their safety and reliability is also becoming critical. On the other hand, hardware transient faults are growing in frequency due to the progressive technology scaling and can lead to failures in DNNs. Existing error-resilience techniques for DNNs often suffer from expensive re-computation, or requires significant implementation effort. In this work, we propose Ranger, a novel, low-cost fault correction technique, which directly rectifies the faulty output due to transient faults without any re-computation. We leverage the insight that DNNs are inherently resilient to benign faults (whose occurrence will not corrupt the program output), and introduce a technique to transform the critical faults (which can result in erroneous output) into benign faults. Ranger is an automated transformation applied to restrict the ranges of values in particular DNN layers, which can dampen the large deviations typically caused by critical faults to smaller ones (the reduced deviation can be tolerated by DNNs). Our evaluation on 8 DNNs (including two used in autonomous vehicles applications) demonstrates that Ranger is highly effective in rectifying the faulty outputs due to transient faults with negligible overheads.