Bezier Crossover Boosts Evolutionary Adversarial Attacks
Researchers have introduced MoCo-EA, an evolutionary strategy that substitutes the conventional crossover mechanism with a Bezier curve operator. Adversarial examples exist on interconnected manifolds, allowing intermediate points to preserve the effectiveness of the attack. This innovative approach enhances transferability while also decreasing both convergence time and the number of queries required.
Key facts
- MoCo-EA introduces a Bezier crossover operator for evolutionary adversarial attacks.
- Adversarial examples exhibit mode connectivity.
- Intermediate points along optimized paths achieve higher transferability than endpoints.
- Bezier crossover outperforms discrete genetic operations.
- The method reduces convergence time and query requirements.
- The paper is available on arXiv with ID 2605.18919.
- The approach does not require gradient information.
- The key insight is that adversarial perturbations lie on connected manifolds.
Entities
Institutions
- arXiv