Volume - 4 | Issue - 3 | september 2022
DOI
10.36548/jaicn.2022.3.004
Published
15 September, 2022
One of the biggest threats in the speaker verification system is that of fake audio attacks. Over the years several detection approaches have been introduced that were designed to provide efficient and spoof-proof data-specific scenarios. However, the speaker verification system is still exposed to fake audio threats. Hence to address this issue, several authors have proposed methodologies to retrain and finetune the input data. The drawback with retraining and fine-tuning is that retraining requires high computation resources and time while fine-tuning results in degradation of performance. Moreover, in certain situations, the previous data becomes unavailable and cannot be accessed immediately. In this paper, we have proposed a solution that detects fake without continual-learning based methods and fake detection without forgetting in order to develop a new model which is capable of detecting spoofing attacks in an incremental fashion. In order to retain original model memory, knowledge distillation loss is introduced. In several scenarios, the distribution of genuine voice is said to be very consistent. In several scenarios, there is consistency in distribution of genuine voice hence a similarity loss is embedded additionally to perform a positive sample alignment. The output of the proposed work indicates an error rate reduction of up to 80% as observed and recorded.
KeywordsDeep learning fake without forgetting data training continuous learning fake audio detection
