SENTRY framework slashes ViT reliability testing costs by 10,700x

Researchers Pramit Kumar Bhaduri, Mahdi Taheri, and colleagues have published SENTRY (Statistical Reliability Analysis of Vision Transformers Under Soft Errors) on arXiv, addressing a critical gap in deploying Vision Transformers for safety-critical applications like autonomous driving and medical imaging. As ViTs grow to massive parameter counts, exhaustive fault injection becomes infeasible. SENTRY leverages finite-population sampling theory to provide formal reliability guarantees, demonstrating that failure rates can be bounded within a 1% margin at 99% confidence using only a few thousand samples—regardless of model scale. This achieves up to a 10,700x reduction in experimental cost compared to exhaustive approaches.

Through extensive evaluation of ViT-Tiny and ViT-Small architectures, the study uncovers a highly non-uniform reliability landscape. While only 3% of FP32 bit-flips result in failure, the vast majority of these events lead to catastrophic accuracy collapse. The researchers localized specific vulnerabilities to normalization layers and critical exponent bits within the IEEE-754 floating-point format. These findings provide a mathematical foundation and actionable insights for designing hardened ViT architectures suitable for edge deployment, ensuring reliability without sacrificing the state-of-the-art accuracy that makes ViTs attractive for safety-critical systems.