Hybrid Machine Learning Models for Deepfake Detection in Multimedia Content

Abstract

The spread of AI-generated synthetic media or so-called deepfakes is not only a significant and growing threat to the integrity of digital information at the social, political, and judicial levels but also a critical one. Current systems of detection with single-modality deep learning models are not sufficiently robust to face unseen manipulation methods and out-of-distribution examples. The paper introduces the Hybrid Deepfake Detection Network (HDDN), which is a single multimodal framework, which combines three complementary branches of analysis: a spatial feature extractor, which relies on an EfficientNet-B4 backbone, a bidirectional Long Short-Term Memory (BiLSTM) component, which analyzes temporal inconsistencies, and a Discrete Wavelet Transform (DWT) frequency- These modules are combined using learned attention-weighted concatenation mechanism that selectively focuses on most discriminative modality according to the input. Extensive tests on FaceForensics++, Deepfake Detection Challenge (DFDC), and Celeb-DF v2 show that Hddn can achieve a detection accuracy of 97.6 percent and an AUC of 0.991 on FaceForensics++, with a cross-dataset AUC drop of just 8.9 percent - and well outperforming all single-modality and two-branch baselines. The applicability of each branch of analysis can be verified by Ablation analysis which confirms the non-redundant contribution of each branch of analysis, and confirms the multimodal hybrid design philosophy. The suggested framework can run at 34 frames per second on a single GPU, which meets the real-time deployment requirements. These results cement hybrid multimodal architectures as the most viable pathway to the next-generation multimedia forensic systems that will be able to keep abreast of the ever-changing environment of generative AI synthesis.