Closed Unkrible closed 2 years ago
Hello, According to Eq(6) in paper, $g = \nabla{x} J ( T(x), y; \phi)$ In attack.py/Line95, $g = \nabla{T(x)} J(T(x), y; \phi)$
attack.py/Line95
Does this changed gradient make a difference?
The changed gradient has little effect on the attack success rate, difference is only about 1%. We do this mainly to speed up backpropagation.
Hello, According to Eq(6) in paper, $g = \nabla{x} J ( T(x), y; \phi)$ In
attack.py/Line95
, $g = \nabla{T(x)} J(T(x), y; \phi)$Does this changed gradient make a difference?