Third-party cyber evaluations involving OpenAI models
openai.com·13h ago
AI Research
arXiv CS.AI·May 21·6 min read
MoCo-EA: Exploiting Adversarial Mode Connectivity for Efficient Evolutionary Attacks
TL;DR
Evolutionary algorithms (population-based search methods) for adversarial attacks—finding inputs that fool AI models without gradient information—traditionally use crossover (blending solutions) that destroys attack effectiveness through discrete interpolation. MoCo-EA replaces standard crossover with Bézier crossover, a continuous curve-based operator that maintains adversarial properties while exploring the solution space.
✦ Why It Matters
Engineers building adversarial robustness tests can use gradient-free attacks that are 2–3× more sample-efficient, critical for black-box security audits.
Key Takeaways
- Replace discrete crossover with BĂ©zier curves to preserve adversarial perturbation properties.
- MoCo-EA reduces function evaluations needed for successful black-box attacks.
- Leverage mode connectivity from loss landscape research to improve evolutionary search.
- Test model robustness efficiently without gradient access or excessive query budgets.
Full Summary
Evolutionary algorithms enable adversarial attacks—crafting inputs that fool machine learning models—without requiring gradient information, making them useful for black-box scenarios. However, traditional crossover operations (combining two solutions) use discrete interpolation that destroys the adversarial properties of candidate perturbations.
Mode Connectivity Evolutionary Attack (MoCo-EA) introduces a BĂ©zier crossover operator that maintains adversarial effectiveness by optimizing perturbations along continuous curves rather than discrete points. This approach leverages mode connectivity, a concept from loss landscape analysis showing that neural network solutions can be connected smoothly.
By replacing discrete interpolation with continuous optimization, MoCo-EA achieves higher attack success rates with fewer function evaluations, reducing computational cost while improving reliability.