Uncovering and Mitigating Destructive Multi-Embedding Attacks in Deepfake Proactive Forensics

Lixin Jia, Haiyang Sun, Zhiqing Guo, Yunfeng Diao, Dan Ma, Gaobo Yang

Introduction

Proactive deepfake forensics has emerged as a promising defense against synthetic media. By embedding invisible "digital fingerprints" or watermarks into original images, researchers aim to track the provenance of content even after it has been manipulated. However, this paper identifies a critical "blind spot" in existing watermarking technologies: Multi-Embedding Attacks (MEA).

Overview of proactive forensics (a) and the destructive impact of Multi-Embedding Attacks (b) on forensic signal integrity.

In real-world scenarios, an image rarely remains untouched. It may be re-watermarked by social media platforms during automated processing or intentionally targeted by malicious actors who embed secondary watermarks to overwrite original forensic evidence. The authors demonstrate that for most state-of-the-art models, a second embedding pass causes the Bit Error Rate (BER) of the original watermark to skyrocket, rendering the forensic tool ineffective.

The MEA Threat and the "Clean Carrier" Fallacy

Current watermarking models are typically optimized under the assumption that the input image is a "clean carrier." When these models encounter an image that already contains a watermark, the encoder often treats the existing signal as redundant texture or noise to be optimized away. This leads to a catastrophic loss of the original forensic information. Experimental results show that without specific defenses, the average BER across prominent models like SepMark and EditGuard jumps from near-zero to over 38% after a single additional embedding.

Adversarial Interference Simulation (AIS)

To mitigate this vulnerability, the authors propose Adversarial Interference Simulation (AIS), a model-agnostic training paradigm designed to enhance the resilience of existing encoder-decoder architectures. AIS introduces three key innovations:

  1. Simulation-in-the-Loop: During the fine-tuning phase, the model is explicitly trained on images that already contain a watermark ($w_1$). The encoder is then tasked with embedding a second watermark ($w_2$), forcing it to handle "non-clean" carriers.
  2. Resilience-Driven Loss ($L_{AIS}$): A specialized loss function penalizes the model if the decoder fails to recover the original watermark ($w_1$) from the doubly-watermarked image.
  3. Sparse Feature Learning: By requiring the recovery of $w_1$ in the presence of $w_2$, AIS encourages the encoder to find "sparse" and decoupled feature dimensions. Instead of spreading the signal across the entire image, the model learns to concentrate forensic information into stable subsets of features that are harder for subsequent embeddings to disrupt.

Experimental Evaluation

The researchers evaluated AIS using the CelebA-HQ and LFW datasets against six baseline frameworks. The results were definitive:

  • Recovery of Integrity: Models enhanced with AIS saw dramatic improvements. For instance, SepMark’s BER dropped from ~48% to a mere 0.04% under MEA.
  • Multi-Round Robustness: AIS-enhanced models maintained low BERs (often under 2%) even after five successive embedding rounds, whereas baseline models failed completely after the second round.
  • Visual Fidelity: Metrics such as PSNR and SSIM confirmed that the increased robustness does not degrade image quality; the watermarks remain imperceptible.

Interestingly, the study found that LampMark—a landmark-based embedding strategy—showed higher natural resilience to MEA because its signal is tied to facial geometry rather than pixel-level textures.

Conclusion

The discovery of Multi-Embedding Attacks exposes a significant vulnerability in current proactive forensic systems. By shifting the training focus from "clean carriers" to "adversarial interference," the AIS paradigm provides a robust, plug-and-play solution that ensures forensic watermarks can survive the complex lifecycle of digital images on the modern web.

Summary updated

Abstract

With the rapid evolution of deepfake technologies and the wide dissemination of digital media, personal privacy is facing increasingly serious security threats. Deepfake proactive forensics, which involves embedding imperceptible watermarks to enable reliable source tracking, serves as a crucial defense against these threats. Although existing methods show strong forensic ability, they rely on an idealized assumption of single watermark embedding, which proves impractical in real-world scenarios. In this paper, we formally define and demonstrate the existence of Multi-Embedding Attacks (MEA) for the first time. When a previously protected image undergoes additional rounds of watermark embedding, the original forensic watermark can be destroyed or removed, rendering the entire proactive forensic mechanism ineffective. To address this vulnerability, we propose a general training paradigm named Adversarial Interference Simulation (AIS). Rather than modifying the network architecture, AIS explicitly simulates MEA scenarios during fine-tuning and introduces a resilience-driven loss function to enforce the learning of sparse and stable watermark representations. Our method enables the model to maintain the ability to extract the original watermark correctly even after a second embedding. Extensive experiments demonstrate that our plug-and-play AIS training paradigm significantly enhances the robustness of various existing methods against MEA.
BibTeX
@inproceedings{jia2026uncovering,
  author = {Lixin Jia and Haiyang Sun and Zhiqing Guo and Yunfeng Diao and Dan Ma and Gaobo Yang},
  title = {{Uncovering and Mitigating Destructive Multi-Embedding Attacks in Deepfake Proactive Forensics}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2026},
  url = {https://openreview.net/forum?id=ImTAKMskfw},
}
Figure