A Hitchhiker's Guide to White-Box Neural Network Watermarking Robustness

Carl De Sousa Trias, Mihai Mitrea, Enzo Tartaglione, Attilio Fiandrotti, Marco Cagnazzo, Sumanta ChaudhuriEUVIP 2023 · DOI

TL;DR

This paper introduces a novel geometric attack called "neuron permutation" for white-box neural network watermarking, demonstrating its effectiveness in disrupting watermark detection without impacting model performance. The authors also propose a counter-attack that successfully restores watermark detectability, even when combined with other common attacks.

What Problem Does This Paper Address?

Neural networks (NNs) are increasingly used in critical applications, making it crucial to protect the intellectual property embedded in their design and to ensure their integrity. Watermarking NNs is a promising solution for this, but existing methods often lack robustness against various attacks. Specifically, the robustness of NN watermarking has primarily been evaluated against "removal attacks" (like pruning or fine-tuning) that aim to degrade the watermark directly. However, the paper highlights a gap in understanding robustness against "geometric attacks," which aim to desynchronize the watermark detector without altering the NN's core functionality. Furthermore, the applicability of NN watermarking beyond image classification tasks has been underexplored. This paper addresses these limitations by formalizing a new geometric attack and its counter-attack, and by evaluating NN watermarking across diverse tasks.

Methods

The core of this paper lies in its contributions to NN watermarking robustness:

  • Formalization of Neuron Permutation Attack: The authors introduce a novel geometric attack that exploits the inherent symmetry in NN layers. By permuting neurons within a layer (and consequently their corresponding input channels in the next layer), the watermark's synchronization with its detector is disrupted. This attack is mathematically described by equations (1) and (2).
  • Development of a Geometric Counter-Attack: To combat the neuron permutation attack, a counter-attack is proposed. This method leverages cosine similarity to identify and correct neuron permutations. By comparing the cosine similarity between the weights of an un-attacked model and a potentially attacked model, the original positions of neurons can be recovered, enabling effective resynchronization of the watermark detector. Equations (3) and (4) outline this approach.
  • Cross-Task Evaluation: The study extends the evaluation of NN watermarking robustness beyond image classification. It investigates the effectiveness of existing watermarking methods on image segmentation and video coding tasks, using state-of-the-art architectures and datasets for each.

Datasets

The paper utilizes standard datasets for its experimental evaluation:

  • Image Classification: CIFAR-10 and Caltech-101.
  • Image Segmentation: Cityscapes.
  • Video Compression: Vimeo-90k (for training) and UVG-dataset (for testing).

Key Insights and Ideas

  • Geometric Attacks are a Real Threat: The paper convincingly demonstrates that neuron permutation is a viable and effective geometric attack against white-box NN watermarking. This attack is particularly insidious because it doesn't degrade the NN's performance, making it hard to detect without specific countermeasures.
  • Existing Methods are Vulnerable: The study reveals that current state-of-the-art white-box watermarking methods are not inherently robust against this new geometric attack. Depending on the specific watermarking technique, either neuron or channel permutations (or both) can completely break watermark detection.
  • Counter-Attack Restores Robustness: The proposed cosine similarity-based counter-attack proves effective in restoring watermark detectability after a neuron permutation attack, even when combined with common removal attacks like Gaussian noise, pruning, fine-tuning, and quantization.
  • NN Watermarking is Versatile: The research successfully extends the evaluation of NN watermarking robustness to tasks beyond image classification, specifically segmentation and video coding. This suggests that NN watermarking is not limited to a single domain and can be applied more broadly.

Key Results

  • The neuron permutation attack successfully disrupts watermark detection (C-BER ≥ 70% and Pearson's coefficient < 0.5) without impacting the NN's performance.
  • The proposed geometric counter-attack effectively synchronizes the watermark detector, achieving high robustness metrics (e.g., C-BER close to 100% or Pearson's coefficient > 0.6) even when the geometric attack is combined with removal attacks.
  • The evaluation across different tasks (classification, segmentation, video coding) shows that existing watermarking methods can be applied to these domains with comparable impacts on performance and robustness, opening doors for broader adoption.

Limitations and Future Work

  • Combined Attack Complexity: While the paper demonstrates the effectiveness of the counter-attack against geometric attacks combined with individual removal attacks, the authors acknowledge that the behavior when multiple removal attacks are combined with geometric attacks requires further in-depth experimental study.
  • Generic Regularization Term: The authors suggest future work on developing a generic regularization term that can be dynamically used as a counter-attack, potentially offering a more unified approach to robustness.
  • Synergies with Cryptographic Attacks: Investigating the coupling of removal, geometric, and cryptographic attacks is identified as a promising avenue for future research to understand the full threat landscape.

Summary updated

BibTeX
@inproceedings{trias2023hitchhiker,
  author = {Carl De Sousa Trias and Mihai Mitrea and Enzo Tartaglione and Attilio Fiandrotti and Marco Cagnazzo and Sumanta Chaudhuri},
  title = {{A Hitchhiker's Guide to White-Box Neural Network Watermarking Robustness}},
  booktitle = {EUVIP},
  year = {2023},
  doi = {10.1109/euvip58404.2023.10323067},
}
Figure