Face Template Protection Through Incremental Learning and Error-Correcting Codes

Face template protection are increasingly applied in identity authentication systems in recent years since it can provide high matching performance. Face template protection aims to provide an effective solution to ensure the security of face features. In most of the end-to-end face template protection schemes, the main challenge of the problem is that the unavailability of the enrolled users’ data makes it difficult to effectively update the model’s parameters. To overcome this challenge, we proposed a KDAT (Knowledge Distillation and Attention Transfer) scheme updating the parameters by incrementally training the model, using new users’ data and public available unlabeled data. The total loss applies a binary cross-entropy loss to learn the new data, and a distillation loss and attention loss to retain the knowledge acquired from the enrolled users. The experimental results on the Extended Yale B, PIE, and FEI datasets show that this scheme can still maintain a high recognition accuracy after multiple incremental training.