Scalable, Privacy-Preserving, and Traceable Multimodal Deepfake Detection in a Cloud-Native Serverless Architecture

Gavhane, Sanjana Raju (2025) Scalable, Privacy-Preserving, and Traceable Multimodal Deepfake Detection in a Cloud-Native Serverless Architecture. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (1MB) | Preview
Preview	PDF (Configuration Manual) Download (2MB) | Preview

Abstract

As AI-generated deepfake content has heated up rapidly, there are highly critical issues of media validity, misinformation, and ethical responsibility. Current methods of detection are usually limited in their scalability, model independence, and weak privacy-preserving procedures, and are therefore more susceptible to failure in the real world, especially in cloud-based settings. The proposed solution in this research is a scalable, automated, and privacy-preserving deepfake detection system that combines a multimodal system, i.e., visual features and audio features to detect inconsistencies in manipulated content. Capsule Networks are used to extract facial features whereas the audio streams get classed to mel spectrograms, and their scores fused to provide better detection performance. The system is created in two phases. During the first phase, the model is trained offline on Google Colab with the FakeAVCeleb dataset. Single face frames are analyzed and visual features are extracted and audio streams turned into mel spectrograms. They are filtered through the Capsule + Score Fusion model with high accuracy and good classification results. The second phase consists of implementing a trained model in the cloud-native, serverless architecture (AWS offerings on S3, Lambda, and DynamoDB). As soon as video is uploaded to the S3 bucket, inference is initiated automatically and content is processed and results saved with related metadata- guaranteeing secure, version-controlled storage and auditability. This two-step scheme achieves both high levels of reliable, near real-time detection with high levels of privacy protections and scalability. The findings suggest the given architecture is technically competent and flexible, which is why it could be incorporated into the large-scale digital content verification-based workflow and media forensics.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Samarawickrama, Yasantha UNSPECIFIED
Subjects:	Q Science > QH Natural history > QH301 Biology > Methods of research. Technique. Experimental biology > Data processing. Bioinformatics > Artificial intelligence Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Artificial intelligence T Technology > T Technology (General) > Information Technology > Cloud computing
Divisions:	School of Computing > Master of Science in Cloud Computing
Depositing User:	Ciara O'Brien
Date Deposited:	26 Mar 2026 09:30
Last Modified:	26 Mar 2026 09:30
URI:	https://norma.ncirl.ie/id/eprint/9215

Actions (login required)

View Item