An AI-Based Gait Authentication Framework Leveraging Swing Phase Dynamics

Gait signals captured by accelerometers embedded in mobile and wearable devices provide an effective and unobtrusive mechanism for user authentication. While deep neural networks have demonstrated strong performance in gait recognition, manually engineered features offer greater interpretability and deeper insight into the structural dynamics of human walking patterns. In this paper, to the best of our knowledge, we are the first to formally establish a direct relationship between accelerometer signals and specific foot movements. Our analysis reveals a distinctive pattern within the swing phase of the gait cycle that carries significantly greater discriminative power than the stance phase. Motivated by this observation, we propose nine novel features extracted exclusively from the swing phase and systematically evaluate their effectiveness for authentication. An authentication system built on these features achieves a 95.51% correct classification rate and a 4.56% equal error rate, outperforming baseline models that rely on full gait-cycle features. Extensive experimental evaluation and detailed computational algorithms further demonstrate the robustness and reliability of the proposed authentication framework.