Custom soft spatial attention mechanism for DeepFake detection using EfficientNet-B7

Abstract

DeepFakes pose significant risks to digital security by enabling realistic facial manipulations that can evade conventional visual inspection. This study presents an attention-enhanced EfficientNet-B7 framework with a Custom Soft Spatial Attention (CSSA) module designed to localize manipulation-sensitive facial regions, including eye boundaries, mouth contours, blending boundaries, and skin-texture discontinuities. Evaluation is conducted on three benchmarks, a balanced Kaggle image subset containing 20,000 facial images, Celeb-DF (v2) with 50,000 sampled frames, and FaceForensics++ Low Quality with 40,000 sampled frames. The proposed model achieves 93.28% held-out test accuracy on the Kaggle subset, representing a 4.94 percentage-point improvement over standard EfficientNet-B7 and a 1.83 percentage-point improvement over an EfficientNet-B7+CBAM variant. Cross-benchmark evaluation yields 87.45% accuracy on Celeb-DF (v2) and 87.31% on FaceForensics++ (LQ). The model achieves a held-out test accuracy of 93.28% and a ROC AUC of 0.97. With 69.8M parameters, 15.0 GFLOPs, a 266 MB model footprint, and 150 ms single-image inference latency under the stated hardware configuration, the framework provides a transparent accuracy-efficiency trade-off for operational screening. Ablation testing confirms a 4.94 percentage-point contribution from CSSA, and LIME-based qualitative analysis supports interpretation of the learned decision regions.