Multimodal biometric systems provide prospective advantages over traditional unimodal systems and are employed diversely for numerous applications. However, issues of data privacy and identity-theft emerges with the extensive use of these systems. To address these issues we propose a multimodal cancelable biometric system based on realtime Deep Feature Unification (DFU). For this, keys-based generic feature extraction is introduced to achieve revocability and dimensionality reduction. Non-invertibility is obtained through random projection of Key Deep features to Query Deep features. Proposed adaptive graph-based fusion process not only extracts complementary information across multiple modalities but also generates multimodal Unified template. The cross diffusion of normalized and optimal graphs ensure the unlinkability and robustness to dynamic environment. Proposed biometric system is assessed over benchmark datasets and shows promising performance against state-of-the-art methods. Average DI and EER achieved by proposed method are 10.35 and 0.12, respectively. Further, robustness against adversary attacks is demonstrated.