Local Conditional Watermarking for Geometrically Robust Deepfake Proactive Forensics

Abstract

This letter proposes LocMark, a local feature modulation-based conditional watermarking method, to address the limitation that existing watermarking methods for deepfake proactive forensics are vulnerable to geometric distortions. Specifically, LocMark establishes the spatial synchronization capability of the conditioned encoder and decoder via local feature modulation. It first utilizes the local mask as the input condition of the encoder to guide a passive localization network, and uses the localization result as the synchronization condition for the decoder, thereby achieving watermark synchronization and blind extraction. Furthermore, considering that deepfake manipulations primarily occur in the face region, the background mask is utilized as the prior condition of the encoder, thus enhancing the watermarking robustness against deepfake distortions. Experimental results demonstrate that, when subjected to individual or composite pixel-value-based and geometric distortions, LocMark can still utilize the synchronized condition mask to effectively extract the watermark message, significantly enhancing the geometric robustness of proactive forensics. Quantitatively, under individual geometric distortions, LocMark achieves an average bit error rate below 1%, whereas the baseline methods approach 50%; meanwhile, under composite deepfake and geometric distortions, LocMark outperforms the two SOTA local watermarking methods by approximately 30%.