The use of hardware redundancy in CubeSats is often limited by physical and budgetary constraints, making alternative approaches to fault tolerance essential for maintaining attitude determination performance. The extended and unscented Kalman filters are typically used to combine all sensor information in a centralized fashion. Federated Kalman filters offer improved fault isolation since each sensor group is associated with an independent local filter, whose estimates are fused by a master filter. Conventional anomaly detection relies on either a Mahalanobis distance- or residual-based measure. These methods require manual threshold selection and do not capture temporal patterns, limiting their effectiveness especially for gradual or subtle faults. In this work, machine learning (ML)-based alternatives are suggested and compared to these conventional approaches, showing a significant increase in detection performance while overcoming some limitations of the traditional methods. The increased computational load associated with these alternatives is assessed against typical microcontroller-based on-board computers used for attitude determination on CubeSats, which was found to be feasible under moderate inference rates. The results demonstrate that the use of ML-based detection within a federated Kalman filter can substantially enhance the reliability of CubeSat attitude determination systems.