Year: 2025
Type: conference-paper
Abstract: As deep neural networks (DNNs) are increasingly deployed in critical applications, ensuring their reliability under hardware faults becomes a key concern. The size of the models and the number of weights in a typical network continually increases. As a consequence, there is a strong pressure to reduce the overhead of adding error correction capabilities to the stored data. This approach involves identifying which bit positions are most critical for the dependability of the DNN. This task, particularly in some novel arithmetic formats like Posit, may require extensive fault injection experiments with high computational costs to assess accuracy degradation. In this paper, the impact of permanent faults on network weights is analyzed. A low-cost analysis based on the faulty/golden ratio of the affected parameter is presented as a tool to predict fault criticality according to the faulty bit position. To validate this methodology, a set of fault injection campaigns was conducted on a typical DNN application for image classification implemented resorting to two very common 32-bit data formats: Posit and Floating Point. The observed accuracy degradation produced by faults injected in different bit positions suggests that the proposed low-cost analysis is a powerful tool to predict bit criticality.
Institutions: Universidad de la República de Uruguay, Politecnico di Torino
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4409561423