I am Shashank V, a final year ECE student in India, an working on a side project on adversarial attacks on Image classification specifically for faces. I found this amazing paper title "Universal adversarial perturbations" in CVPR 2017 and wanted to try the code, which the authors so generously provided on GitHub(https://github.com/LTS4/universal).
Since they have trained their model on ImageNet and the most standard dataset for faces in the VGG-face, I had to compute my own adversarial template for VGG-face dataset. While doing so, I found that the variable 'v' which outputs this adversarial template had all its values as NaN.
Furthermore, the main reason for this was when I found that the function in the deepFool code which is utilized
dr = project_boundary_polyhedron(ddf,ff(idx),Q);
outputs a zero value which results in this NaN.
I have tried different image datatypes to try solving this issue and also tried different parameters but in vain.
I was wondering if anyone could kindly help me on this, as I guess I have reached a dead-end and any advice would really be appreciated.
Thanks!