Compressed deepfake detection via GA-LASSO selection of deep features and machine learning models.

Area of Science:

• Computer Science
• Artificial Intelligence
• Digital Forensics

Background:

• Deepfake detection faces challenges from high-dimensional data, compression, and poor generalization.
• Existing methods struggle with real-world media conditions.

Purpose of the Study:

• To propose a hybrid feature-selection framework for robust and efficient deepfake detection.
• To reduce feature dimensionality while preserving detection accuracy.

Main Methods:

• A hybrid framework combining genetic algorithms (GA) and LASSO regularization.
• Feature selection applied to ResNet50 embeddings, reducing dimensions from 2048 to 120-170.
• Experiments conducted on FaceForensics++ (FF++) and Celeb-DF v2 datasets with varying compression levels (C0, C23, C40).

Main Results:

• Achieved high accuracy (e.g., AUC = 99.48%, 97.11% accuracy on FF++ C23) and maintained competitiveness in cross-dataset and cross-manipulation tests.
• Demonstrated robust performance under harsh compression (e.g., 78.74% AUC on Celeb-DF v2 C40 with SVM).
• GA+LASSO significantly reduced computational cost compared to GA alone, especially under heavy compression.

Conclusions:

• The proposed GA+LASSO framework offers a lightweight and robust solution for deepfake detection.
• It enhances accuracy, generalization, and stability while reducing computational load.
• The method is well-suited for real-world media analysis due to its efficiency and resilience.