Threshold Attack remark

[aliotopal]

Threshold Attack is not suitable for high-resolution inputs such as 224 x 224 x 3 ImageNet images because it uses evolutionary strategies to search across all image pixels, leading to the creation of extremely large arrays. For example, it attempts to allocate an array with shape (150,528, 150,528), which requires more than 169 GB of memory. This greatly exceeds the memory capacity of most standard systems, which typically have between 8 and 64 GB of RAM, making the attack impractical for large-scale models. Additionally, the ART implementation does not offer configuration options to reduce this memory demand.