Fingerprint authentication is a widely used method for secure identity verification due to its uniqueness and reliability. With the growing demand for biometric systems in various applications, fingerprint-based authentication systems need to improve their performance and privacy. Traditional Fingerprint Recognition (FPR) approaches often encounter several challenges. These include limitations in feature extraction accuracy, inefficiency in processing large-scale data, and significant concerns regarding the privacy and security of stored biometric information. The key gap in current solutions lies in the lack of integration between lightweight neural network architectures and effective feature selection mechanisms, simultaneously ensuring the security of fingerprint data. Most existing systems either focus on improving recognition performance or enhancing data privacy, but not both in a unified framework. This study introduces a novel SaDeXNet, a deep learning model combining Depthwise Separable Convolutions with Spatial Attention to enhance fingerprint feature representation. SaDeXNet improves recognition accuracy while maintaining computational efficiency. Additionally, fingerprint features are secured to address privacy concerns using the Cheon Kim Kim Song (CKKS) Fully Homomorphic Encryption (FHE) scheme. The performance is evaluated using the Equal Error Rate (EER), demonstrating the integration of the spatial attention boost feature quality, and the use of CKKS FHE protects the fingerprints. This dual approach ensures a secure, accurate, and efficient fingerprint authentication system.