Deepfakes threaten the trustworthiness of online media and public discourse. We present a lightweight yet effective hybrid for facial deepfake detection that couples MobileNet spatial encoders with sequence-level LSTM, probabilistic Hidden Markov Models (HMM) for temporal regularization, and Kolmogorov–Arnold Networks (KAN) for adaptive, interpretable nonlinear modeling. On FaceForensics++ and Celeb-DF v2—two widely used and challenging benchmarks—our best variant achieves strong cross-validated accuracy, outperforming both our baselines and previously reported methods on these datasets. We report precision, recall, F1-score, and AUC, and discuss training choices that yield robust generalization under video compression. Our results suggest that combining LSTM-based sequence modeling with HMM smoothing and KAN adapters is a practical path toward reliable, low-latency deepfake screening in real-world pipelines.