DeepDistAL: Deepfake Dataset Distillation Using Active Learning

Abstract

In the rapidly evolving landscape of artificial intelligence (AI), particularly in the Deepfake domain, largescale datasets play a pivotal role in ensuring performance, including the model’s accuracy, robustness, trustworthiness, etc. However, the increasing size and intricacy of the datasets impose a growing demand for computational resources and amplify the cost and duration of model building. To mitigate the challenge, dataset distillation provides a solution. For the Deepfake detection problem, noteworthy datasets such as VDFD, FaceForensics++, DFDC, and Celeb-DF underscore the indispensability of extensive data for ensuring model robustness. Nevertheless, the computational requirement associated with these datasets presents significant obstacles. This paper describes a data distillation method utilizing Active Learning to reduce dataset size while retaining essential data qualities. The proposed method facilitates efficient model training selecting representative samples by capturing the most salient features, thereby enabling effective performance in resource-constrained environments. The study encompasses developing a data distillation algorithm tailored for Deepfake detection, rigorous experimentation with a major Deepfake dataset to validate its efficacy, and a comprehensive comparison of the model performance trained on distilled versus original datasets. Through thorough analysis, we demonstrate the practicality and effectiveness of our proposed method in alleviating computational demands without compromising detection accuracy.