Neural Network-Based Deepfake Detection

This thesis focuses on deepfake detection, the detection of an advanced synthetic media technology that generates deceptively authentic yet forged image videos. The proposed SparcoNet (Spatial Cost-efficient Neural network) achieves an average of 0.985 AUC among six state-of-the-art deepfake datasets. The network performance is further improved with self-supervised learning, which reduces the success rate of adversarial attacks by up to 85% and improves the inter-cross data evaluation performance by an average of 12%. The defense ability of the model is enhanced against black-box adversarial attacks by introducing Block Switching, which reduces the attack success rate by more than 90%.