@techreport{oai:ipsj.ixsq.nii.ac.jp:00214328,
 author = {Hideaki, Takahashi and Kohei, Ichikawa and Keichi, Takahashi and Hideaki, Takahashi and Kohei, Ichikawa and Keichi, Takahashi},
 issue = {10},
 month = {Dec},
 note = {Federated learning is a distributed learning method in which multiple clients cooperate to train a model. Each client sends the gradient of its locally trained model to the server, and the server aggregates the received gradients to build a global model. Since federated learning requires many clients to train a high-performance model, researchers have designed incentive mechanisms that distribute rewards to clients to motivate their participation. While most incentive mechanisms distribute rewards according to the contribution of each client often defined by the number of data, little research has been done on the risk that clients try to claim more rewards by overstating the number of data. This paper proposes three possible methods to exaggerate the size of a local dataset: simple exaggeration of the reported number, modification of the batch size during training, and exaggeration of the dataset by Data Augmentation. Using a variety of models and datasets, we show the inadequacy of current anomaly detection methods in identifying such exaggerations., Federated learning is a distributed learning method in which multiple clients cooperate to train a model. Each client sends the gradient of its locally trained model to the server, and the server aggregates the received gradients to build a global model. Since federated learning requires many clients to train a high-performance model, researchers have designed incentive mechanisms that distribute rewards to clients to motivate their participation. While most incentive mechanisms distribute rewards according to the contribution of each client often defined by the number of data, little research has been done on the risk that clients try to claim more rewards by overstating the number of data. This paper proposes three possible methods to exaggerate the size of a local dataset: simple exaggeration of the reported number, modification of the batch size during training, and exaggeration of the dataset by Data Augmentation. Using a variety of models and datasets, we show the inadequacy of current anomaly detection methods in identifying such exaggerations.},
 title = {Difficulty of detecting overstated dataset size in Federated Learning},
 year = {2021}
}