A Brief Review for Compression and Transfer Learning Techniques in DeepFake Detection

View on arXiv ← Back to list

Authors: Andreas Karathanasis, John Violos, Ioannis Kompatsiaris, Symeon Papadopoulos

Published: 2025-04-29 13:37:21+00:00

AI Summary

This paper investigates compression and transfer learning techniques for deepfake detection on edge devices. It evaluates pruning, knowledge distillation, quantization, fine-tuning, and adapter-based methods across Synthbuster, RAISE, and ForenSynths datasets, demonstrating effectiveness even at 90% compression when training and testing data share the same deepfake model, but revealing a domain generalization problem otherwise.

Abstract

Training and deploying deepfake detection models on edge devices offers the advantage of maintaining data privacy and confidentiality by processing it close to its source. However, this approach is constrained by the limited computational and memory resources available at the edge. To address this challenge, we explore compression techniques to reduce computational demands and inference time, alongside transfer learning methods to minimize training overhead. Using the Synthbuster, RAISE, and ForenSynths datasets, we evaluate the effectiveness of pruning, knowledge distillation (KD), quantization, fine-tuning, and adapter-based techniques. Our experimental results demonstrate that both compression and transfer learning can be effectively achieved, even with a high compression level of 90%, remaining at the same performance level when the training and validation data originate from the same DeepFake model. However, when the testing dataset is generated by DeepFake models not present in the training set, a domain generalization issue becomes evident.

Key findings

High compression levels (up to 90%) are achievable with minimal accuracy loss when training and testing data come from the same deepfake generator. However, performance significantly degrades when the test data originates from different generators, highlighting a domain generalization challenge. Knowledge distillation generally outperforms pruning, and adapters enhance performance.

Approach

The research explores combining model compression (pruning, knowledge distillation, quantization) and transfer learning (fine-tuning, adapter-based methods) to reduce the computational demands of deepfake detection models while maintaining accuracy. They evaluate these techniques on various datasets and compression levels.

Datasets

Synthbuster, RAISE, ForenSynths, and a 'dogs vs. cats' dataset for transfer learning.

Model(s)

A VGG-based model (approximately 4.5 million parameters).

Author countries

Greece

← Previous