Controlled Model Debiasing through Minimal and Interpretable Updates