ENTROPY INTEGRATED DYNAMIC ROUTING IN CAPSULE NETWORKS AND DEEPFAKE DETECTION APPLICATIONS

Abstract

Capsule networks are recognized for their ability to capture part-whole hierarchies in visual data, but they face challenges in managing uncertainty within part-object relationships. This study introduces an entropy-integrated dynamic routing algorithm to enhance both the performance and interpretability of capsule networks by embedding information-theoretic principles. Our method incorporates an entropy-based regularization term in the final iteration of the routing process, improving routing decisions and reducing reliance on uncertain capsule connections. Evaluations on CIFAR-10 demonstrate a mean accuracy of 85.58%, surpassing the baseline accuracy of 84.61% achieved with standard dynamic routing. Comparative analysis with recent entropy-based routing approaches highlights our method’s balance of computational efficiency and routing flexibility, achieved without additional model complexity. These findings position entropy-integrated dynamic routing as a powerful tool for enhancing the interpretability and effectiveness of capsule networks in high-uncertainty environments. Our study finds application in deepfake detection by embedding entropy into the capsule networks’ dynamic routing process. Traditional capsule networks capture spatial hierarchies critical for identifying synthetic media, yet their scalability is limited by uncertainties in iterative routing. To address this, we put into practice our entropy-based regularization that adjusts routing coefficients based on capsule activation uncertainty, selectively prioritizing reliable connections and enhancing interpretability and model robustness without significant added complexity. Experiments on FaceForensics++ and DeepPhy datasets reveal that our entropy-integrated routing algorithm achieves higher classification accuracy in both binary and multi-class deepfake detection tasks, outperforming baseline models like CapsuleForensics. These results underscore entropy regularization’s potential to improve capsule networks’ capacity to address nuanced spatial inconsistencies in manipulated media, advancing scalable and interpretable solutions in deepfake detection.