Attention-augmented hybrid framework with evolutionary optimization for robust deepfake detection

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Generative models like GANs and Autoencoders have increased deepfake production, threatening digital trust and enabling fraud.
• Existing deepfake detection models lack generalizability and struggle with temporal feature modeling across diverse datasets and formats.
• A research gap exists in developing robust and adaptable deepfake detection systems.

Purpose of the Study:

• To design a novel deepfake detection framework with high accuracy and generalizability.
• To enhance temporal inconsistency and facial dynamic capture in manipulated videos.
• To create a scalable and adaptive deepfake detection solution.

Main Methods:

• A multi-stage pipeline utilizing Vision Transformers (ViTs) for spatial features and pretrained CNNs (MobileNetV2) for feature refinement.
• A hybrid Gated Recurrent Unit (GRU) architecture with a deep-fake-specific gating mechanism for enhanced sequential learning.
• Genetic Algorithm optimization for GRU parameters, layer count, and hyperparameter tuning.

Main Results:

• The proposed model significantly outperforms state-of-the-art methods in accuracy, precision, recall, F1-score, and ROC-AUC.
• Demonstrated strong generalization capabilities on unseen data, reducing false positive rates.
• Achieved superior performance on benchmark datasets like Celeb-DF V2 and FaceForensics++.

Conclusions:

• The novel GRU-based framework with evolutionary optimization offers an effective and scalable approach to deepfake detection.
• The system shows significant promise for applications in digital forensics, content verification, and policy enforcement.
• This research addresses the need for robust deepfake detection against evolving manipulation techniques.