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

Abstract

In an era of advanced synthetic media, deepfake detection is challenged by high-dimensional feature spaces, compression artifacts, and poor generalization. This paper proposes a hybrid feature-selection framework combining genetic algorithms (GA) with LASSO regularization to reduce redundancy in ResNet50 embeddings from 2048 to 120-170 features (>90% reduction). Experiments on FaceForensics++ (FF++) and Celeb-DF v2 under C0, C23, and C40 compression show improved accuracy, efficiency, and robustness. In single-seed evaluations, the method achieves AUC = 99.48 and 97.11% accuracy (KNN, Deepfakes C23) and remains competitive under cross-dataset and cross-manipulation testing. On Celeb-DF v2 with harsh C40 compression, SVM achieves AUC = 78.74, outperforming many end-to-end models. Multi-seed analysis shows consistent top-tier performance across datasets (e.g., FF++ C40: 85.2% accuracy; Celeb-DF C0: 88.7%). GA+LASSO maintains accuracy comparable to GA while substantially reducing computational cost, particularly under heavy compression (Celeb-DF C40). Overall, the proposed framework enhances accuracy, generalization, and stability while reducing feature dimensionality and computational cost, offering a lightweight and robust deepfake detection solution suited to real-world media conditions.