Model-Independent Real-Time Anomaly Detection at CMS Using Knowledge Disillation with CICADA

In the CMS trigger system, CICADA (Calorimeter Image Convolutional Anomaly Detection Algorithm) leverages unsupervised deep learning techniques to enable physics-model independent trigger decisions, making it sensitive to anomalous events. To meet the latency and size requirements, it is trained with KD (knowledge distillation) to compress the model size dramatically while maintaining high performance. The final model is implemented in FPGAs and has recently been deployed at CMS Run 3 at 40 MHz. In this talk, we present the architecture, implementation and performance studies of this new anomaly detection algorithm.