Lightweight CNN Models for Real-Time Fake Image Detection on Mobile Devices

The proliferation of deepfake technology and AI-generated synthetic media is bringing forth extreme threats to digital content authenticity, extending to advanced detection systems deployable on limited-resource mobile devices [1][2][3]. This work presents a systematic investigation of lightweight convolutional neural network (CNN) architectures designed for real-time deepfake detection on mobile devices [4][5][6]. We contrast various optimization strategies like quantization, pruning, and knowledge distillation to attain significant model compression with preserved detection accuracy [7][8][9]. Our Binary Neural Network (BNN) approach attains 20× computational complexity reduction with minimal accuracy loss and attains 92.1% detection accuracy with as few as 1.2M parameters [5][6][10]. Experimental evaluations on widely used benchmark datasets like FaceForensics++, DFDC, and CelebDF show that optimized lightweight models are capable of attaining real-time inference speeds of 12-35ms on mobile devices while maintaining cross-dataset generalization capabilities [1][11][12]. The addition of frequency domain features based on Fast Fourier Transform (FFT) and Local Binary Patterns (LBP) enhances detection performance by 3-5% across various deepfake generation approaches [5][13][6].