The rise of deepfake technology poses significant risks to digital integrity, privacy, and public trust. While visual based detection methods are often vulnerable to compression and visual artifacts, physiology based cues such as remote PhotoPlethysmoGraphy (rPPG) offer a promising alternative due to their reduced susceptibility to such distortions. In this work, we propose a novel deepfake detection framework that exploits physiological consistency through Time-Frequency (TF) analysis. Specifically, facial Regions of Interest (ROIs) are extracted from video frames, and green channel based rPPG signals are computed. These signals are then transformed using the Short-Time Fourier Transform (STFT) to generate phase based spectrograms, which are classified using a lightweight Convolutional Neural Network (CNN). The incorporation of STFT phase information enhances robustness by capturing subtle temporal inconsistencies introduced by generative models. Our method achieves over 99% accuracy on the VidTIMIT vs. DFTIMIT dataset and demonstrates strong performance on other benchmark datasets such as CelebDF and DFDC-P. The results highlight the generalizability and effectiveness of our physiology driven, time-frequency approach for robust deepfake detection.