In an era where digital authenticity is frequently compromised by sophisticated synthetic audio technologies, ensuring the integrity of digital media is crucial. This paper addresses the critical challenges of catastrophic forgetting and incremental learning within the domain of audio deepfake detection. We introduce a novel methodology that synergistically combines the discriminative feature extraction capabilities of SincNet with the computational efficiency of LightCNN. Our approach is further augmented by integrating Feature Distillation and Dynamic Class Rebalancing, enhancing the model’s adaptability across evolving deepfake threats while maintaining high accuracy on previously encountered data. The models were tested using the ASVspoof 2015, ASVspoof 2019, and FoR datasets, demonstrating significant improvements in detecting audio deepfakes with reduced computational overhead. Our results illustrate that the proposed model not only effectively counters the issue of catastrophic forgetting but also exhibits superior adaptability through dynamic class rebalancing and feature distillation techniques.