CheckOne: Efficient Fault Detection for Vision Transformers

A new research paper introduces CheckOne, a lightweight method for detecting and mitigating hardware faults in Vision Transformers (ViTs), which are increasingly used in safety-critical applications. The paper, titled 'CheckOne: Lightweight Fault Detection and Mitigation for Vision Transformers,' is available on arXiv under the identifier 2608.04035. The authors comprehensively evaluate ViT reliability, highlighting the need for symmetric protection across layers. CheckOne significantly reduces computational overhead compared to conventional Algorithm-Based Fault Tolerance (ABFT) methods, which are challenging for ViTs due to their high computational demands. Experiments with multiple ViTs show that CheckOne mitigates critical faults by up to 26 times and achieves an average 3.8x higher performance than ABFT. The paper is categorized under Computer Science > Hardware Architecture and includes submission history, references, and tools for citation. It also mentions arXivLabs, a framework for collaborative projects that adhere to arXiv's values of openness, community, excellence, and user data privacy.