Assuming signal is real and of length n, n even, then
Y(1) is for frequency 0, the DC contribution, and it's real. Don't mess with that point.
Y(2) and Y(n) are complex conjugates. You can multiply one of that pair by exp(i*theta) and the other by exp(-i*theta), where theta is a random angle with 0 <= theta < 2*pi. the new Y(2) and Y(n) remain complex conugates.
In general from k = 2 to n/2, Y(k) and Y(n+2-k) form a complex conjugate pair. For each of those pairs, do the same kind of multiplcation as above, with a different random angle. Each pair remain complex conjugates.
Y(n/2+1) is real. Don't mess with that point either.
Here the random phases are totally uncorrelated from frequency to frequency, which may or may not be physically realistic.