Ensuring the integrity of results in online learning and assessment tools is a challenge, due to the lack of direct supervision increasing the risk of fraud. We propose and evaluate a machine learning-based method for detecting anomalous behaviour in an online retrieval practice task, using an XGBoost classifier trained on keystroke dynamics and task performance features to distinguish between genuine and fraudulent responses. The classifier requires only a modest amount of training data--approximately 100 short-answer responses, typically collected within 10 minutes of practice-- and maintains good performance when not all feature types are available. This method enhances the reliability of online learning and assessment by identifying anomalous response behaviour in a way that preserves learners' privacy. [For the complete proceedings, see ED675583.]