Facial Expression Recognition (FER) serves as a core task in emotion analysis and human–computer interaction. Despite major progress achieved through deep learning, the heavy computational cost and model size of current methods restrict their suitability for real-time or embedded applications. Standard fine-tuning methods require updating millions of weights, whereas recent Parameter-Efficient Fine-Tuning (PEFT) techniques adapt only a small subset of parameters. However, the purpose and relevance of normalization strategies within these methods remain underexplored for FER, in particular, convolutional architectures. In this work we explore Weight-Decomposed Low-Rank-Adaptation (DoRA) using Nuclear norm regularization as an alternative normalization strategy. Results from our experiments showed that nuclear norm DoRA consistently improved adaptation efficiency and accuracy over existing PEFT methods while maintaining a reduced parameter footprint. These results help in laying a foundation for efficient and deployable FER systems and also highlight the wider potential of normalization-aware fine-tuning in both vision and language domains. Source code: https://github.com/saisyam-1729/Efficient-FER .