Deepfake technology poses a serious threat to the authenticity of digital media. To proactively counter malicious manipulation at the source, this work introduces an active defense framework that integrates adaptive watermark embedding with blockchain-based notarization, enabling strong image authentication and end-to-end traceability. A U-Net-based embedding network equipped with an Adaptive Embedding Module is developed to seamlessly embed watermark signals into facial images while minimizing perceptual distortion. The embedded watermark is simultaneously registered on the blockchain to ensure immutability. To enhance visual quality and resistance to detection bypass, an adversarial-example-driven enhancement strategy is employed to suppress embedding artifacts. A dedicated extraction network is then designed to accurately recover the embedded watermarks under various deepfake scenarios, exhibiting strong robustness. Finally, recovered watermarks are cross-verified against blockchain records, forming a transparent and trustworthy evidentiary chain for image forensics. Experimental results demonstrate that the proposed framework significantly improves both detection robustness and forensic traceability, while preserving high visual fidelity of the protected images.