In this study, we propose a novel local SAR prediction framework based on deep learning. To this end, we introduce multi-task feedback adversarial learning to simultaneously predict local SAR distribution and its peak SAR value. The proposed model learns a mapping between simulated B1+ magnitude/tissue property maps and local SAR provided by EM simulations. Given query inputs with the properly trained model, the generator produces the local SAR distribution slice by slice, and the local SAR peak estimator predicts the upper bound of local SAR values. Validation results show that the proposed model may allow online subject-specific local SAR prediction.