Third-party cyber evaluations involving OpenAI models
openai.com·14h ago
AI Research
arXiv CS.AI·Jun 9·5 min read
SENTRY: Statistical Reliability Analysis of Vision Transformers Under Soft Errors
TL;DR
As Vision Transformers (ViTs) are increasingly used in critical applications like autonomous driving, their reliability against soft errors (temporary faults) is essential. A statistical fault injection framework was developed using finite-population sampling theory to assess ViTs' reliability.
✦ Why It Matters
Engineers can use this framework to evaluate and enhance the reliability of Vision Transformers in critical applications.
Key Takeaways
- Adopt SENTRY for efficient reliability analysis of ViTs.
- Focus on vulnerabilities in normalization layers during testing.
- Utilize finite-population sampling for cost-effective fault injection.
- Ensure failure rates are within 1% at 99% confidence.
How It Works
SENTRY employs finite-population sampling theory to statistically analyze the reliability of Vision Transformers. By simulating soft errors through a limited number of fault injections, it provides formal guarantees on failure rates, significantly reducing the need for exhaustive testing.