This paper proposes a novel method to address the hyperparameter sensitivity issue of existing machine unlearning methods. While existing methods utilize the weight vector of a single model fine-tuned with specific hyperparameters, this paper utilizes the weight vectors of multiple models trained with various hyperparameters. Machine unlearning is performed by collecting only elements with consistent signs among multiple vectors to generate a new vector and subtracting it from the weight vector of the original model. Experimental results using 12 datasets and 4 backbone architectures show that the proposed method outperforms existing state-of-the-art methods while consuming similar or less computational resources.
