attack arXiv Oct 30, 2025 · Oct 2025
Zhichao Hou, Weizhi Gao, Xiaorui Liu · North Carolina State University
Efficient iterative adversarial attacks via selective layer activation recomputation, matching full-budget adversarial training at 30% cost
Input Manipulation Attack vision
This work tackles a critical challenge in AI safety research under limited compute: given a fixed computation budget, how can one maximize the strength of iterative adversarial attacks? Coarsely reducing the number of attack iterations lowers cost but substantially weakens effectiveness. To fulfill the attainable attack efficacy within a constrained budget, we propose a fine-grained control mechanism that selectively recomputes layer activations across both iteration-wise and layer-wise levels. Extensive experiments show that our method consistently outperforms existing baselines at equal cost. Moreover, when integrated into adversarial training, it attains comparable performance with only 30% of the original budget.
cnn transformer North Carolina State University
ML Security Papers