Inconsistent privacy budget

[Drxdx]

📚 Documentation

When I run the normal DPSGD, with a privacy budget of 1, the noise assumption is 1.2. However, when I used the DISK optimizer and set the "noise" to 1.2, the final privacy budget was not equal to 1. Has anyone encountered this problem?

[Drxdx]

c = (1 - kappa) / (gamma * kappa)
norm_factor = math.sqrt(c**2 + (1 - c) ** 2)
noise_multiplier = noise_multiplier / norm_factor.
If I need to keep the privacy budget consistent, do I need to multiply by this coefficient? Is this correct?

[iden-kalemaj]

Adding @564612540 to this thread since they are the author of DISK paper. Can you also share what your final privacy budget was? Going from budget to estimated noise and then back from noise to budget might not give the exact same value since the estimated noise is computed to achieve the desired budget approximately.