Lightweight Multimedia Anomaly and Integrity Detection for Consumer IoT Using Knowledge Distillation

The rapid growth of Consumer Internet of Things (CIoT) devices has significantly increased real-time multimedia data exchange, heightening vulnerability to attacks targeting audio, video, and image content. This paper introduces Multimedia Anomaly and Integrity Detection using Knowledge Distillation (MAID-KD), a lightweight multi-task framework that performs anomaly detection and integrity verification in CIoT environments. MAID-KD leverages a Transformer-based teacher model to extract rich spatio-temporal features from multimedia streams, while a compact CNN-LSTM student model optimized for edge deployment is trained through feature alignment, soft-target distillation, and variational projection. Experimental results demonstrate that MAID-KD achieves superior accuracy and F1-score compared to state-of-the-art baselines, while reducing model size and inference latency by over 60%. These results highlight MAID-KD's ability to deliver scalable, privacy-preserving, and multimedia-aware security for CIoT devices such as smart surveillance systems, health-monitoring wearables, and connected home platforms.