<p>Anomaly detection in system logs is critical for monitoring modern software systems and identifying abnormal behaviors associated with reliability incidents and security violations. While dynamic analysis and deep learning-based detectors have achieved strong performance, they remain vulnerable to adversarial perturbations that manipulate observable log sequences through execution-level behavior morphing. To address this challenge, we propose SENTRY, an adversarially robust log-sequence anomaly detection framework that integrates shapelet-based pattern unit extraction, confidence-guided time-step masking, and teacher–student knowledge distillation with difficult-sample learning. Specifically, we enhance the shapelet algorithm with a customized matching-based distance metric to extract discriminative pattern units from categorical event-ID sequences, and reconstruct log sequences to suppress adversarial noise. The reconstructed sequences are then processed by a sequence model equipped with a confidence-guided masking mechanism that reweights each time step according to its normality confidence, thereby emphasizing anomaly-relevant evidence under camouflage attacks. Finally, SENTRY improves robustness via a difficulty-aware distillation strategy that guides the student model to learn from hard and misclassified cases, stabilizing decision behavior under diverse attacks. Extensive experiments on three real-world log datasets demonstrate that SENTRY consistently outperforms strong baselines on both clean and adversarial inputs, and additional studies on alternative backbones further validate the generalizability of the proposed framework.</p>