Fragile Model Watermark for integrity protection: leveraging boundary volatility and sensitive sample-pairing

Zhenzhe Gao, Zhenjun Tang, Zhaoxia Yin, Baoyuan Wu, Yue Lu

Abstract

Neural networks have increasingly influenced people's lives. Ensuring the faithful deployment of neural networks as designed by their model owners is crucial, as they may be susceptible to various malicious or unintentional modifications, such as backdooring and poisoning attacks. Fragile model watermarks aim to prevent unexpected tampering that could lead DNN models to make incorrect decisions. They ensure the detection of any tampering with the model as sensitively as possible. However, prior watermarking methods suffered from inefficient sample generation and insufficient sensitivity, limiting their practical applicability. Our approach employs a sample-pairing technique, placing the model boundaries between pairs of samples, while simultaneously maximizing logits. This ensures that the model's decision results of sensitive samples change as much as possible and the Top-1 labels easily alter regardless of the direction it moves.

Introduction

As Deep Neural Networks (DNNs) are integrated into critical infrastructure, maintaining their integrity is a significant security concern. Models can be subtly altered post-deployment through malicious "backdoor" injections or unintentional modifications like pruning for optimization. Unlike robust watermarks intended to survive such changes, fragile watermarks are designed to fail upon any modification, serving as a high-sensitivity alarm.

The primary challenges in fragile watermarking include achieving high sensitivity to minute changes, ensuring computational efficiency during sample generation, and maintaining universality across different types of tampering.

Methodology

The proposed method generates "sensitive sample pairs" through a three-stage process designed for both efficiency and sensitivity.

The three-stage workflow for generating sensitive sample pairs, involving binary layer simplification, logit maximization, and gradient-based pairing.
  1. Stage 0: Binary Classification Layer: To manage high-dimensional multi-class models, a temporary, user-specific binary classification layer is added. This simplifies the boundary problem and allows for user-specific watermarking.
  2. Stage 1: Logit Maximization: Samples are optimized to maximize raw output scores (logits) while remaining near the decision boundary. This ensures that the neurons involved are highly active and central to the model's learned features.
  3. Stage 2: Sample-Pairing via Gradient Ascent: Using small-step gradient ascent, the algorithm creates a pair of samples: one just inside and one just outside the decision boundary.

Key Insights: Boundary Volatility

The core innovation lies in leveraging boundary volatility. Regions near decision boundaries are inherently unstable; even minor weight adjustments caused by tampering are likely to shift these boundaries.

Conceptual illustration of how model integrity attacks shift the original classification boundary, causing the classification of sensitive sample pairs to change.

By placing a sample pair precisely on either side of this "fence," any shift causes both samples to fall on the same side, altering their relative classification and triggering a detection alert. Maximizing neuron activation further ensures these samples are tied to core model structures, enhancing sensitivity to even the most subtle fine-tuning (e.g., learning rates as low as $1e-9$).

Experimental Evaluation

The method was rigorously tested on three standard benchmarks:

  • CIFAR-10 (10 classes)
  • GTSRB (43 traffic sign classes)
  • Flowers102 (102 fine-grained categories)

Evaluations spanned various architectures, including ResNet18, VGG16, and ResNet152, demonstrating the approach's versatility across different model depths and complexities.

Results

  • High Detection Accuracy: The method achieved a 92% to 99% detection rate for backdoor attacks, significantly outperforming baseline methods like DBI and AID.
  • Extreme Sensitivity: It successfully detected modifications even when the tampering learning rate was so low that standard performance metrics remained unchanged.
  • Robustness to Pruning: The watermark effectively detected weight pruning at rates as low as 0.1%.
  • Efficiency: By focusing on output logits rather than internal neuron iteration, the generation process is significantly faster than previous state-of-the-art methods, making it viable for large-scale models.

Conclusion

This fragile watermarking technique provides a robust and efficient solution for DNN integrity protection. By combining boundary volatility with sensitive sample-pairing, it offers unprecedented sensitivity to unauthorized modifications. Future work will explore enhancing performance against quantization and investigating defenses against watermark-aware adversaries.

Summary updated

Abstract

Neural networks have increasingly influenced people's lives. Ensuring the faithful deployment of neural networks as designed by their model owners is crucial, as they may be susceptible to various malicious or unintentional modifications, such as backdooring and poisoning attacks. Fragile model watermarks aim to prevent unexpected tampering that could lead DNN models to make incorrect decisions. They ensure the detection of any tampering with the model as sensitively as possible.However, prior watermarking methods suffered from inefficient sample generation and insufficient sensitivity, limiting their practical applicability. Our approach employs a sample-pairing technique, placing the model boundaries between pairs of samples, while simultaneously maximizing logits. This ensures that the model's decision results of sensitive samples change as much as possible and the Top-1 labels easily alter regardless of the direction it moves.
BibTeX
@inproceedings{gao2024fragile,
  author = {Zhenzhe Gao and Zhenjun Tang and Zhaoxia Yin and Baoyuan Wu and Yue Lu},
  title = {{Fragile Model Watermark for integrity protection: leveraging boundary volatility and sensitive sample-pairing}},
  booktitle = {ICME},
  year = {2024},
  doi = {10.1109/icme57554.2024.10688355},
}
Figure