Biometric authentication techniques are crucial for ensuring the security and reliability of identification processes. Traditional biometric systems relying on fingerprint, face, or iris recognition face challenges such as spoofing and hygiene concerns. However, palm print recognition offers a non-intrusive, highly accurate, and robust solution to combat counterfeiting. This paper presents a novel algorithm for palm print recognition using multi-spectral scans, with the goal of mapping palm images to a more compact euclidean space. We present a new approach to enhance the learning capabilities of the model by updating the triplets during the training process, resulting in improved triplet formation. Our approach utilizes a deep convolutional network known as EfficientNet-B1 in combination with global local self-attention module which is trained to optimize inherently the embeddings. We achieved the state-of-the-art recognition results by employing a 15-way 2-shot learning. Extending this achievement further proposed model’s effectiveness was tested in adverse situations where the training and testing occur on separate datasets. For this, additional experiments were conducted using various datasets namely, CASIA multispectral, IIT Delhi palmprint, and TONGJI dataset.