Relative L2 sensitivity assumption allows for adaptive noise levels in the Relative Gaussian Mechanism, ensuring privacy in gradient descent.