Existing research on privacy-preserving facial recognition has primarily focused on securing biometric templates, but it performs similarity distance computations in plain text, exposing sensitive data during the matching process. Additionally, most approaches lack secure encrypted storage and comprehensive data protection mechanisms throughout the facial recognition workflow, leaving biometric data vulnerable during storage, transmission, and processing. To address these limitations, this research proposes a privacy-preserving facial recognition framework that performs direct cosine distance computation on encrypted facial feature vectors using Fully Homomorphic Encryption. By calculating the L2 norm before encryption, the system enables similarity matching entirely within the encrypted domain, thereby protecting data confidentiality during processing. The encrypted feature vectors are securely stored and managed in a decentralized, tamper-resistant Hyperledger Fabric blockchain and the InterPlanetary File System (IPFS), ensuring data integrity and secure management throughout the facial recognition workflow. Experimental evaluations conducted on the Labeled Faces in the Wild (LFW) dataset show that the proposed system achieves superior performance with 85\% accuracy in facial recognition using cosine distance in ciphertext, significantly surpassing the 73.33\% accuracy obtained through Euclidean matching in plaintext. The system maintains efficient execution times on a blockchain ledger containing 60 encrypted entries, with registration completed in 1.8761 seconds and encrypted verification requiring only 4.2489 seconds. This framework ensures that sensitive biometric information remains encrypted and protected during processing, storage, and transmission, effectively minimizing the risk of information leakage while supporting the development of secure and efficient biometric authentication solutions for modern facial recognition applications.