基于FakeDetect-MobileNet模型的虚假图像检测研究OA

The rapid advancement of deepfake technology presents dual challenges of accuracy and com-putational efficiency for deepfake image detection.To address the high computational complexity of tradi-tional deep learning models and their difficulty in real-time application on mobile devices,a lightweight detection model FakeDetect-MobileNet was proposed,based on the MobileNetV3 architecture.This model employed a two-stage training strategy:Stage 1 to train only the classification layer by freezing the pre-trained feature extraction layers and;Stage 2 to perform fine-tuning of the entire network.Overfitting was mitigated through data augmentation and multiple regularization techniques,improving the model's gener-alization capability.Experimental results on the Kaggle Deepfake Detection dataset demonstrate FakeDe-tect-MobileNet achieve a detection accuracy of 97.34%,with only 3.99 million parameters and a CPU-based inference speed of 11 FPS.Compared to mainstream models,it reduce parameter count by 25.7%and 50.7%compared with EfficientNetB0 and DenseNet121,respectively,while reducing inference la-tency by 31%and 58%.By maintaining high detection accuracy while significantly reducing resource consumption,this model is particularly suitable for mobile device deployment,providing a practical solu-tion for real-time applications such as social media content moderation and news image verification.