Abstract
The rapid progress of deepfake generation technologies poses a growing threat to the credibility of digital video content. Most existing solutions rely on reactive detection methods, which struggle to generalize against continuously evolving generative models. To address this challenge, this paper proposes a proactive and verifiable deepfake mitigation framework that secures video content at the point of creation. The proposed system integrates three complementary components: (i) a spatio-temporal attention model based on EfficientNetV2B0 and a two-layer Long Short-Term Memory (LSTM) network to identify perceptually and temporally salient regions, (ii) an attention-guided Generative Adversarial Network (GAN) to embed an imperceptible yet robust watermark, and (iii) a blockchain-based ledger to provide immutable file-level integrity verification via cryptographic hashing. Experimental results show that the attention model achieves 97.20% accuracy, confirming its ability to learn discriminative spatio-temporal representations. The watermarking scheme preserves high visual quality, achieving an average PSNR of 33.67 dB and SSIM of 0.9838. The proposed framework detects 100% of face-swapping attacks generated using DeepFaceLab 2.0 and demonstrates robustness against common video degradations, including recompression (90%), scaling (89%), and blurring (98%). Performance degradation under severe Gaussian noise (80%) highlights the practical limits of the watermark under non-structural pixel perturbations. An ablation study further confirms that both attention guidance and temporal modeling are critical to achieving high robustness. Overall, this work shifts deepfake defense from reactive detection to proactive, content-aware, and verifiable protection.