In this paper, we propose a novel approach to overcome the limitations of existing unlearning methods by selectively “forgetting” certain data by applying “patches” to the neural network. While existing methods are expensive and difficult to verify because they retrain the model with the remaining data, our method guarantees data deletion by finding and applying a lightweight minimal patch. To delete multiple data points or entire classes, we use a method to selectively and iteratively delete representative data points. Experimental results on several datasets show that the proposed method is efficient and consumes less memory while maintaining model performance.
