FedReverse: Multiparty Reversible Deep Neural Network Watermarking

Junlong Mao, Huiyi Tang, Yi Zhang, Fengxia Liu, Hongliang He, Shanxiang Lyu2024 · DOI

TL;DR

This paper introduces FedReverse, a novel method for embedding watermarks into Deep Neural Networks (DNNs) that allows multiple parties to collaboratively protect their intellectual property after the model is trained. Crucially, FedReverse is reversible, meaning watermarks can be completely removed with consent, and it's designed to be robust against common attacks, all while minimally impacting model accuracy.

What Problem Does This Paper Address?

The increasing prevalence and complexity of Deep Neural Networks (DNNs) have made protecting the intellectual property (IP) embedded within them a significant challenge. As models become more valuable, unauthorized use, plagiarism, and lack of attribution are growing concerns. Traditional methods for protecting DNNs, such as watermarking, often have drawbacks: they might require access to the model during training, negatively impact performance, or be difficult to remove. Furthermore, in collaborative settings like Federated Learning (FL), where multiple parties contribute to a single model, ensuring individual IP claims and preventing watermark conflicts among participants is particularly tricky. This paper aims to address these issues by proposing a multiparty, reversible watermarking solution that works post-training and is secure.

Methods

FedReverse builds upon the concept of reversible watermarking by difference contraction. Here's a breakdown of its key features:

  • Post-Training Embedding: Unlike many methods that embed watermarks during the training phase, FedReverse embeds them into the model's weights after training is complete. This is crucial for collaborative scenarios where clients might not want to reveal their training data or process.
  • Multiparty Collaboration: FedReverse allows multiple clients to embed their unique watermarks into the same model. This is achieved through an orthogonal key generation technique. Each client is assigned a unique key (a vector $u_i$ and a dither $d_i$). The embedding process projects the model's weights onto these orthogonal vectors, effectively creating separate "spaces" for each client's watermark.
  • Reversibility: The core of FedReverse is its reversibility. By using a difference contraction method, the original model weights can be perfectly recovered after the watermark is removed. This requires the unanimous consent of all involved parties to ensure that no single party can unilaterally remove a watermark.
  • Security: FedReverse is designed to be resistant to Known Original Attacks (KOA). This means attackers who have access to original (unwatermarked) and watermarked versions of the model will find it very difficult to infer the secret keys or forge watermarks. The orthogonal key generation and the nature of the embedding process contribute to this security.
  • Difference Contraction: This technique, adapted from reversible data hiding, is used to embed information into floating-point numbers (DNN weights). It involves quantizing the data and then contracting the difference between the original and quantized values, adding a controlled amount of "noise" that allows for perfect reconstruction.

Datasets

The paper uses the MNIST dataset for training and evaluating its proposed method. This dataset consists of 60,000 training and 10,000 testing grayscale images of handwritten digits.

Key Insights and Ideas

  • Post-training watermarking is feasible and beneficial: The ability to watermark after training opens up new possibilities for IP protection, especially in collaborative environments like Federated Learning, without interfering with the initial training process.
  • Orthogonality is key for multiparty watermarking: Using orthogonal vectors for client keys is a clever way to ensure that watermarks embedded by different clients don't interfere with each other, preventing conflicts and allowing for independent extraction.
  • Reversibility with unanimous consent is a strong security/usability feature: This allows for flexibility. A model can be watermarked for IP protection, and then the watermark can be removed if the model is sold or shared under specific conditions, requiring agreement from all parties who contributed to the watermark.
  • Robustness against KOA is achievable: The method demonstrates that even with knowledge of original and watermarked pairs, inferring the secret keys or forging watermarks is computationally challenging due to the proposed key generation and embedding strategy.

Key Results

  • Minimal Accuracy Impact: FedReverse shows a negligible impact on model accuracy. For example, in a scenario with 20 clients, the test accuracy for FedReverse was 0.9918, very close to the original model's accuracy of 0.9920, and superior to other methods like FedTracker (0.9833) and WAFFLE (0.9829).
  • Reversibility Demonstrated: The paper visually shows that watermarked weights can be recovered to their original state after the watermark removal process.
  • Robustness: The theoretical analysis and simulations suggest FedReverse is robust against various attacks, particularly Known Original Attacks (KOA).
  • Scalability: The experiments show that increasing the number of clients has a relatively limited impact on model accuracy, indicating good scalability. For instance, accuracy remained around 0.99 for CNN and 0.94 for MLP even with an increased number of clients.
  • Parameter Sensitivity: The study highlights that parameters like $\alpha$ and $\Delta$ influence accuracy and distortion. Higher values generally lead to increased distortion (higher MSE, lower SWR) and a slight decrease in accuracy, suggesting a trade-off that needs careful tuning.

Limitations and Future Work

  • Unanimous Consent for Removal: While a security feature, requiring unanimous consent for watermark removal could be a practical bottleneck if one party is unavailable or uncooperative.
  • Impact of $\alpha$ and $\Delta$: The paper notes that higher values of $\alpha$ and $\Delta$ can lead to increased distortion and a decline in accuracy. Finding optimal ranges for these parameters is crucial for balancing watermark strength and model performance.
  • Future Work: The paper suggests that further research could explore more advanced key generation techniques and investigate the impact of FedReverse on different types of DNN architectures and datasets. The authors also imply that optimizing the trade-off between watermark robustness, reversibility, and model performance remains an ongoing area of interest.

Summary updated

Abstract

The rising complexity and cost of training DNN models underscore the need for intellectual property protection. In this regard, DNN watermarking has emerged as a crucial safeguarding technique. This paper introduces FedReverse, a multiparty reversible watermarking method that ensures robust copyright protection and minimal performance impact. FedReverse is reversible, allowing collaborative watermark embedding post-training and enables complete removal with unanimous consent. It is resistant to Known Original Attacks (KOA), making watermark forgery and key inference difficult. Comprehensive simulations with MLP and CNN models on varying parameters show FedReverse’s robustness, reversibility, and minimal accuracy impact.
BibTeX
@misc{mao2024fedreverse,
  author = {Junlong Mao and Huiyi Tang and Yi Zhang and Fengxia Liu and Hongliang He and Shanxiang Lyu},
  title = {{FedReverse: Multiparty Reversible Deep Neural Network Watermarking}},
  year = {2024},
  doi = {10.1109/hpcc64274.2024.00213},
}
Figure