LightKD-Detector: An Empirical Investigation into Lightweight Deepfake Detection via Knowledge Distillation under Computational Constraints

Nhaloor, Govind (2025) LightKD-Detector: An Empirical Investigation into Lightweight Deepfake Detection via Knowledge Distillation under Computational Constraints. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (820kB) | Preview
Preview	PDF (Configuration Manual) Download (440kB) | Preview

Abstract

The introduction of Generative Adversarial Networks (GANs) in 2014, and the consequent rise of Diffusion Models in 2020, hyper-realism of deepfakes has spread, becoming a threat to the reliability and wholesomeness of information in society. Despite hardware improvements between the years 2018 and 2025, deepfake generation continues to outperform the detection capabilities on consumer devices. As of 2025, the major model for detection of deepfakes are run on server grade hardware with strong computing power and are not deployable on consumer level hardware where misinformation is spread rapidly. There is an immediate need to eradicate the trade-off between the size of the model, detection accuracy and the speed of inference. The proposed research relies on LightKD-Detector, a model that utilizes Knowledge Distillation (KD) to load knowledge of a heavy model (EfficientNet-B4) onto a lightweight model (MobileNetV2). The models had been trained on the FaceForensics++ dataset and then evaluated on the more challenging Celeb-DF (v2) dataset at a consumer grade GPU (RTX 3060) to replicate potential resource constraints in the real world. The intended efficiency was exceeded with the proposed MobileNetV2 Student model, achieving an impressive inference rate and a light model. However, the model collapsed in terms of mode collapse in its detections. These results are crucial indicators of the "Resource Gap" of AI forensics. As shown in the paper, though lightweight architectures resolve the issue of latency, the success of Knowledge Distillation is inevitably tied to the convergence of the Teacher model that is not possible in resource-constrained systems without pre-trained weights. This experimentally validates the fact that the entry barrier to strong deepfake defense is still low because of its complexity.

Item Type:	Thesis (Masters)
Supervisors:	Name Email McCabe, Liam UNSPECIFIED
Uncontrolled Keywords:	Deepfake Detection; Knowledge Distillation; MobileNetV2; AI Forensics; Resource Constraints; Generative AI
Subjects:	Q Science > QA Mathematics > Computer software > Computer Security T Technology > T Technology (General) > Information Technology > Computer software > Computer Security Q Science > QH Natural history > QH301 Biology > Methods of research. Technique. Experimental biology > Data processing. Bioinformatics > Artificial intelligence > Generative artificial intelligence Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Artificial intelligence > Generative artificial intelligence
Divisions:	School of Computing > Master of Science in Cyber Security
Depositing User:	Ciara O'Brien
Date Deposited:	04 Sep 2026 08:45
Last Modified:	04 Sep 2026 08:45
URI:	https://norma.ncirl.ie/id/eprint/9813

Actions (login required)

View Item