Federated Unlearning for Samples Based on Adaptive Gradient Ascent of Angles

With the development of communication networks, user demand for privacy protection has been growing, and the distributed data structure also brings corresponding challenges. As awareness of the ‘the right to be forgotten’ in privacy protection increases, federated unlearning has become an important solution. Current research mainly focuses on the unlearning of specific classes, but traditional training is typically based on the order of instances. Consequently, precise and efficient unlearning of random samples is required. To address this issue, we present a novel GA-AL scheme that can quickly and effectively achieve the unlearning of random samples. Firstly, when a unlearning request is made, the gradients of all target samples are calculated. The model is then updated in reverse, and the angle of the gradient ascent is adaptively adjusted based on the current gradient direction. Subsequently, lifelong learning is incorporated to make the model more adaptable to the remaining samples, minimizing performance loss. Finally, experiments are conducted on three different types of datasets and models, focusing on four aspects: unlearning effect, efficiency, remaining data performance, and privacy protection. Compared to the Retrain baseline, the GAAL scheme has a performance gap of less than 10% and achieve a 130 times acceleration. This performed best of all experimental methods.