continual learning strategies for anomaly detection

This thesis investigates the application of continual learning to anomaly detection, aiming to develop models that can adapt to new data while preserving previously acquired knowledge. A major challenge is catastrophic forgetting — when a model loses information from earlier training phases — and the related risk of learning anomalies as normal patterns if updates are not properly managed. Through a review of recent research, this work examines state-of-the-art continual learning strategies — including replay, generative replay, compressed replay, and regularisation — and analyses how they balance adaptability and stability. The findings show that replay remains the most effective mechanism for retaining knowledge, while regularisation improves stability and privacy. Compression further enhances scalability, making continual learning feasible for real-world applications.