AmpDet: Dynamical Autoencoding Amplification Detection Against Stealthy Evasion Attacks on Gait Identity Recognition System

Extensive investigations have revealed that stealthy evasion attacks pose severe exogenous threats to gait identity recognition systems, potentially inducing the security collapse of smart wearables. To address this challenge, this article formulates a dynamic autoencoding framework to adapt a novel amplification detection (AmpDet) mechanism, by which the imperceptible noisy perturbations are significantly amplified to the detectable level within a transformed space. Specifically, we formalize a rigorous theorem with a sufficient condition of amplification projection onto the minimal-eigenvalue eigenvector direction to establish a dynamical encoding matrix that can guarantee theoretical amplification feasibility and intensity with lower–upper bounds. Further, we have fully integrated four types of characteristics with invertibility, orthogonality, minimal eigenvalue variability, and dynamical randomization to develop a tractable encoding matrix construction algorithm. Finally, we conduct comparative experiments to validate the superiority of our AmpDet against eight types of noisy perturbations. Across four types of evaluation metrics, AmpDet outperforms six baseline anomaly detectors, including$\chi ^{2}$, CuSum, one-class support vector machine (OC-SVM), IsoFst, long–short-term memory autoencoder (LSTM-AE), and transformer-based autoencoder (Trans-AE) in detecting imperceptible perturbations. Empirical results shed light on the potential universality of AmpDet in safeguarding gait recognizers and mitigating stealthy evasion attacks for a broader range of pattern recognition tasks. The mentioned datasets and demo codes are available athttps://github.com/bigsky2020