Biometric template security is a critical challenge, as compromised templates are irrevocable. This paper proposes a novel hybrid dynamic projection scheme for generating secure and cancelable fingerprint templates. Our method utilizes a ResNet-50 backbone with an ArcFace head to extract a robust 512-dimensional embedding. This embedding is then secured using a unique projection matrix, Rdyn, which is a weighted combination of a biometric-driven component (Rbio) and a user-key-driven random component (Rkey). We evaluated our system on the CASIA-FingerprintV5 dataset. The results demonstrate that our protected template achieves an Equal Error Rate (EER) of 0.467%, a minimal performance loss compared to the 0.435% EER of the raw embeddings. Furthermore, the protected system maintains a Genuine Acceptance Rate (GAR) of 98.9% at a 0.1% False Acceptance Rate (FAR), proving its effectiveness in balancing high-level security with practical accuracy.