A Survey of Adversarial Efficiency Degradation on Vision Transformers with Input-Adaptive Optimization

A survey studies adversarial efficiency degradation attacks targeting input-adaptive mechanisms such as Vision Transformer (ViT) token pruning and early stopping; such attacks can increase model computation without necessarily reducing accuracy. The survey uniformly compares SlowFormer (universal adversarial patch) and DeSparsify (per-image perturbation) on three token pruning frameworks, A-ViT, ATS and AdaViT, reporting in a standardized manner with GFLOPs, accuracy loss and Attack Success (AS) metrics. The survey also reviews whether existing defenses can effectively restore efficiency under attack; low-power deployments such as mobile and embedded make such threats particularly critical.