Federated Learning of User Verification Models Without Sharing Embeddings
Authors: Hossein Hosseini, Hyunsin Park, Sungrack Yun, Christos Louizos, Joseph Soriaga, Max Welling
ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We present the experimental results for user veriļ¬cation with voice, face, and handwriting data and show that Fed UV is on par with existing approaches, while not sharing the embeddings with other users or the server. |
| Researcher Affiliation | Industry | 1Qualcomm AI Research, an initiative of Qualcomm Technologies, Inc. |
| Pseudocode | Yes | Algorithm 1 (Mc Mahan et al., 2017a) Fed Avg. and Algorithm 2 Federated User Authentication (Fed UV). |
| Open Source Code | No | The paper does not contain an explicit statement about releasing its source code or a link to a code repository for the methodology described. |
| Open Datasets | Yes | We present the experimental results for voice, face and handwriting recognition using Vox Celeb (Nagrani et al., 2017), Celeb A (Liu et al., 2015) and MNIST-UV datasets, respectively, where MNIST-UV is a dataset we created from images of the EMNIST dataset (Cohen et al., 2017). |
| Dataset Splits | Yes | We selected 1, 000 speakers and generated 25 training, 10 validation and 10 test examples for each speaker. (Vox Celeb) We selected 1, 000 identities from those who had at least 30 images, which we split into 20, 5 and 5 examples for training, validation, and test sets, respectively. (Celeb A) We created MNIST-UV dataset that contains data of 1, 000 writers each with 50 training, 15 validation, and 15 test examples. (MNIST-UV) |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., CPU/GPU models, memory specifications) used for running its experiments. |
| Software Dependencies | No | The paper does not provide specific software dependency details with version numbers (e.g., Python, PyTorch, TensorFlow versions or other libraries). |
| Experiment Setup | Yes | We train the UV models using the Fed Avg method with 1 local epoch and 20, 000 rounds with 0.01 of users selected at each round. Table 2 provides the network architectures used for each dataset. In models, we use Group Normalization (GN) instead of batch-normalization (BN) following the observations that BN does not work well in non-iid data setting of federated learning (Hsieh et al., 2019). Models are trained with SGD optimizer with learning rate of 0.1 and learning rate decay of 0.01. |