Engineering Reports Инженерия (общая)
Lightweight Ensemble Transfer Learning for Frame‐Level Deepfake Identification
Аннотация
ABSTRACT Advances in generative artificial intelligence have enabled the creation of highly detailed deepfake videos, raising serious concerns about distortion, security, and privacy. Traditional deepfake detection models perform poorly when faced with unseen data distributions and compression conditions. To address these challenges, a lightweight transfer‐learning RGB ensemble framework has been developed for robust detection of AI‐generated faces in video frames, and its cost‐accuracy trade‐off is evaluated against larger detectors. This framework combines heterogeneous lightweight convolutional architectures such as MobileNetV3, ShuffleNetV2, and EfficientNet‐B0 as ImageNet‐pretrained feature extractors and fuses their predictions through equal‐weight soft voting; frame scores are then averaged per video. Evaluation uses 1009 videos (15,878 face‐cropped images), a video‐disjoint 70:15:15 split, and three random seeds (42, 77, and 123) with a new split for each seed. A random frame‐level 80:20 split gives 99.18% accuracy, but every evaluation frame comes from a video that is also in the training set; on held‐out videos the ensemble reaches 94.92% frame‐level and 97.14% ± 3.05 video‐level accuracy (ROC‐AUC 0.998). The ensemble is 1.4% and 2.2% points less accurate than Xception (98.54%) and ViT‐Small (99.38%) but uses 3.1‐ and 3.2‐times fewer parameters and 7.8‐ and 7.3‐times fewer FLOPs; it improves ShuffleNet V2 and EfficientNet‐B0 but not significantly on MobileNetV3 alone (96.92%). The findings are processed under a single dataset and latency was measured on a GPU and a CPU, not on an edge device.