AdaEA: Adaptive Model Ensemble Adversarial Attack for Improving Adversarial Transferability
An Adaptive Model Ensemble Adversarial Attack for Boosting Adversarial Transferability
AI summary
The researchers propose an adaptive model ensemble adversarial attack, AdaEA, which improves the transfer of adversarial examples between models with large architectural differences, such as from CNN to ViT. The method adaptively controls output fusion by monitoring the disparity ratio of each surrogate model's contribution to the adversarial objective, and introduces a disparity-reducing filter to synchronously update the direction. On multiple datasets, AdaEA improves markedly over existing ensemble attacks and can also be stacked onto existing transfer-based attacks for further enhancement.