All in One: Unifying Deepfake Detection, Tampering Localization, and Source Tracing with a Robust Landmark-Identity Watermark

2026-03-02 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy · Depth: Expert, extended

Summary

The LIDMark framework introduces a unified proactive forensics solution for deepfake detection, tampering localization, and source tracing, addressing limitations of existing methods that treat these as independent tasks. This framework utilizes a novel 152-dimensional landmark-identity watermark, termed LIDMark, which combines a 136-D tamper-sensitive facial landmark vector with a 16-D robust source identifier. A Factorized-Head Decoder (FHD) is designed to robustly extract this composite watermark, factorizing shared backbone features into specialized regression and classification heads. The regression head enables an "intrinsic-extrinsic" consistency check for detection and localization, comparing recovered intrinsic landmarks with re-detected extrinsic landmarks. The classification head decodes the source identifier for tracing. Experiments on CelebA-HQ and LFW datasets demonstrate that LIDMark provides a robust, imperceptible, and unified solution, outperforming baselines in visual quality and forensic task performance, even under severe deepfake manipulations.

Key takeaway

For Computer Vision Engineers developing deepfake countermeasures, the LIDMark framework offers a significant advancement by unifying detection, localization, and tracing into a single model. You should consider integrating this trifunctional approach to move beyond reactive forensics, leveraging its high-capacity 152-D watermark and Factorized-Head Decoder for more robust and comprehensive content authentication. This can streamline your forensic pipeline and enhance the granularity of your deepfake analysis.

Key insights

LIDMark unifies deepfake detection, localization, and tracing using a composite watermark and factorized decoder.

Principles

• Combine tamper-sensitive and robust signals in one watermark.
• Factorize decoder heads for heterogeneous payload recovery.
• Utilize intrinsic-extrinsic consistency for detection and localization.

Method

The LIDMark framework embeds a 152-D landmark-identity watermark via a two-stream encoder. A Factorized-Head Decoder (FHD) recovers landmarks for detection/localization via an "intrinsic-extrinsic" consistency check and an identifier for tracing.

In practice

• Embed 152-D LIDMark for comprehensive deepfake forensics.
• Use Average Euclidean Distance (AED) for tamper detection.
• Employ Bit Error Rate (BER) for source identifier recovery.

Topics

• Deepfake Detection
• Tampering Localization
• Source Tracing
• Digital Watermarking
• Factorized-Head Decoder

Code references

• vpsg-research/LIDMark

Best for: Computer Vision Engineer, Research Scientist, AI Researcher, AI Scientist, Deep Learning Engineer

Related on AIssential

• Detecting Deepfakes with Multivariate Soft Blending and CLIP-based Image-Text Alignment — Multimodal image-text alignment and blended data augmentation significantly improve deepfake detection generalization and robustness.
• Where, What, Why: Toward Explainable 3D-GS Watermarking — A representation-native 3D-GS watermarking framework decouples carrier selection and visual compensation for robust, imperceptible, and explainable embedding.
• [P] VeridisQuo - open-source deepfake detector that combines spatial + frequency analysis and shows you where the face was manipulated — Combining spatial and frequency domain analysis enhances deepfake detection, especially for high-quality fakes.
• CUPID: Reconstructing UV Texture Maps for Interpretable Person-of-Interest Deepfake Detection — CUPID uses UV texture maps and MAE for robust, interpretable POI deepfake detection without deepfake training data.
• YouTube expands deepfake detection to wider group of users — YouTube expanded deepfake detection to protect individuals from AI-driven impersonation, using face scans and IDs.
• Architecture-Adaptive Uncertainty Fusion for Deepfake Detection — Correlation-Optimized Fusion adaptively combines five uncertainty sources to improve deepfake detection reliability, especially under distribution shift.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.