This article tackles the ill-posed challenge of solving a nonlinear heat transfer equation in backward form. Traditional regularization methods (Tikhonov, quasi-reversibility, quasi-boundary value, truncation, etc.) exist, but we introduce a novel approach based on nonlocal diffusion equations with a rescaled convolution kernel. We will provide error estimates and demonstrate the accuracy of this technique under specific assumptions.