AIRA-D FusionNet: A Multi-Modal Deep Learning Framework for Violence Recognition through Audio-Visual Cues

This paper presents AIRA-D FusionNet, a multi-modal system for violence recognition that integrates visual analysis using a MoViNet-A0 backbone with auditory processing of MFCC features through a BiLSTM network. Trained on a combined dataset, our fused model achieves a recall of 0.91 and an AUC of 0.856, demonstrating a 12% improvement in recall over unimodal baselines by effectively leveraging complementary audio-visual cues. To enable practical deployment, the model was successfully optimized for mobile inference by conversion to TensorFlow Lite. This confirms the system's viability for real-time violence detection applications on resource-constrained devices, offering a sensitive and efficient solution for automated security monitoring.