Adaptive-LwF: continual training of morphing attack detector without forgetting

In Biometrics, the presence of privacy restrictions on personal data transfer and storage poses significant challenges in creating a sufficiently comprehensive and varied dataset by leveraging various data sources for traditional batch-based training procedures. This is particularly true in the Morphing Attack Detection (MAD) task, in which data involves facial images and a limited number of public datasets of well-controlled images are available. In this context, MAD systems generally suffer from limited generalization capabilities, with low performance on new and unseen data. Therefore, in this paper, we propose Adaptive-LwF, adopting the recent paradigm of Continual Learning (CL) as a viable solution to enable incremental training across multiple sites. Indeed, CL assumes that once a model has been trained, previous data cannot be utilized in subsequent training iterations and can be deleted. In particular, we investigate the performance of different methods in this new scenario, where a model is updated each time a new chunk of data, of variable size, becomes available. We focus our attention on the well-known Learning without Forgetting (LwF) algorithm, proposing a novel adaptive approach able to automatically fine-tune its parameters in relation to the variable size of the specific input chunks. Experimental results confirm that our approach is capable of mitigating the catastrophic forgetting effects, and the superior performance of the Adaptive-LwF algorithm with respect to alternative solutions.