Deception detection is essential for establishing a secure and reliable society. It plays a critical role in high-stakes circumstances like court trials or police interrogations, when the result is greatly impacted by the behaviour of the individual being investigated. Facial expressions can serve as valuable cues for detecting deception, which is a challenging task due to the ability of deceptive individuals to mask their true feelings or intentions. Besides, facial landmarks, essential locations on the face utilized to identify facial muscle movement and shape changes, play an important role in determining deceptive behaviour. Facial landmark detection is crucial in several applications, including facial analysis for 3D reconstruction and expression. It provides a substantial amount of information about an individual. This paper addresses 3D facial landmark extraction, feature selection, and a Bi-directional Long Short-Term Memory (Bi-LSTM) model. Firstly, the Mediapipe framework extracted 478 facial landmarks with coordinates (x, y, z) to enhance the reliability and precision of deception. Next, the 3D facial landmarks most relevant to deception cues were selected using a feature selection-based Pearson correlation coefficient method. Then, 3D facial landmarks are used to train a Bi-LSTM model that predicts deception from truth-tellers using a high-stakes real-life trial dataset. Finally, the model successfully achieved an accuracy of 97%, surpassing other state-of-the-art methods.