MaxiMals: A Low-cost Hardening Technique for Large Vision Transformers
We propose MaxiMals, an experimentally tuned low-cost mitigation solution to reduce the impact of transient faults on Vision Transformers. MaxiMals can correct 73.06% of critical failures on average, with overheads as low as 0.25%.
Authors
- Olivier Sentieys (ORCID: https://orcid.org/0000-0003-4334-6418)
- Fernando Fernandes dos Santos (ORCID: https://orcid.org/0000-0002-3504-9862)
- Paolo Rech (ORCID: https://orcid.org/0000-0002-9597-1007)
- Angeliki Kritikakou (ORCID: https://orcid.org/0000-0002-9293-469X)
- Marcello Traiola (ORCID: https://orcid.org/0000-0003-1484-5162)
- Lucas Roquet (ORCID: https://orcid.org/0009-0001-8458-516X)
Institutions
- University of Trento (IT)
- Institut de Recherche en Informatique et Systèmes Aléatoires (FR)
- Université de Rennes (FR)
Publication Details
- Published
- 2024-09-16
- DOI
- https://doi.org/10.1109/radecs61970.2024.11298581
- Primary Topic
- Industrial Vision Systems and Defect Detection
- Type
- article
- Field-Weighted Citation Impact
- 0.00