Misclassification-driven Fingerprinting for DNNs Using Frequency-aware GANs

Weixing Liu, Shenghua ZhongIJCAI 2025 · DOI

TL;DR

This paper introduces a novel method for protecting the intellectual property (IP) of deep neural networks (DNNs) by generating unique "fingerprints" using frequency-domain perturbations. Their approach leverages a frequency-aware Generative Adversarial Network (GAN) to create misclassified samples, which are then used to verify model ownership against potential theft, achieving state-of-the-art results.

What Problem Does This Paper Address?

Deep neural networks (DNNs) are incredibly valuable assets, but their high development costs make them targets for intellectual property theft. Existing methods to protect DNNs, such as watermarking and fingerprinting, have limitations. Watermarking can sometimes degrade model performance, while fingerprinting often requires extensive training of additional models or generates fingerprints near decision boundaries, potentially compromising image quality and robustness. This paper aims to address the need for an efficient, imperceptible, and robust DNN fingerprinting method that doesn't require significant additional model training.

Methods

The core of this paper is a novel fingerprinting approach that utilizes a frequency-aware Generative Adversarial Network (GAN). Here's a breakdown of their method:

  • Misclassified Samples as Fingerprints: The fundamental idea is to use samples that a DNN misclassifies as its unique fingerprint. For high-performing models, naturally occurring misclassified samples are scarce.
  • Frequency-Aware GAN for Fingerprint Generation: To overcome the scarcity, they propose a GAN that generates adversarial perturbations. These perturbations are applied in the frequency domain of normal input samples. The GAN is designed to:
  • Generator (G): Creates perturbations that, when added to clean samples, cause the source model (MD) to misclassify them.
  • Discriminator (D): Distinguishes between clean and perturbed (misclassified) samples. This discriminator is enhanced with a Fourier Spectrum Discriminator (Dspec) that analyzes frequency-domain features, making it more robust to spectral variations.
  • Frequency Domain Perturbations: The key innovation lies in applying perturbations in the frequency domain. This allows for changes in the image's representation that influence the classifier's decision without causing noticeable visual artifacts in the image itself, thus ensuring imperceptibility.
  • Fingerprint Verification: Once fingerprints are generated, they are used to verify a suspect model. The matching rate between the predicted labels of the source model and the suspect model on these fingerprint samples is calculated. If this rate exceeds a predefined threshold, the suspect model is flagged as stolen.

Datasets

The authors evaluated their method on the following datasets:

  • CIFAR-10: A standard dataset for image classification.
  • GTSRB (German Traffic Sign Recognition Benchmark): A dataset focused on traffic sign recognition.
  • Tiny-ImageNet: A more challenging dataset with a larger number of classes and higher resolution images.

They used a split of these datasets into "defender" and "attacker" sets to simulate real-world IP attack scenarios.

Key Insights and Ideas

  • Frequency Domain is Key for Imperceptibility and Robustness: The paper highlights that manipulating images in the frequency domain is a powerful way to alter their representation for a DNN without introducing visible changes. This is crucial for creating imperceptible fingerprints that are still effective.
  • GANs as Efficient Fingerprint Generators: By using a GAN, the authors can generate an "unconstrained number" of misclassified samples, overcoming the scarcity problem of naturally occurring misclassifications. This makes the fingerprinting process more efficient.
  • Frequency-Aware Discriminator for Enhanced Detection: Incorporating a spectral discriminator in the GAN's discriminator improves its ability to detect subtle frequency-based perturbations, leading to more robust fingerprinting.
  • Misclassified Samples are Inherently Informative: The core idea that misclassified samples reveal unique characteristics of a model's decision-making process is leveraged effectively.

Key Results

The proposed method demonstrated strong performance across various IP threats and datasets:

  • Tiny-ImageNet: Achieved a state-of-the-art (SOTA) AUC of 0.98 under IP removal attacks, outperforming existing methods by 8%.
  • GTSRB: Consistently achieved the best Average Boundary Precision (ABP) for three types of IP detection and erasure attacks.
  • Robustness: The method showed high robustness against various IP removal and detection/evasion attacks, often outperforming other leading fingerprinting techniques.

Limitations and Future Work

  • Limitations: While the paper demonstrates strong performance, the authors acknowledge that in the CIFAR-10 dataset, their method's ABP was moderate compared to some others. They also note that for more complex datasets, the classification boundaries between models vary greatly, which their method handles well, but for simpler datasets, the boundaries are more similar, potentially limiting the impact of misclassified samples.
  • Future Work: The authors plan to extend their frequency-domain perturbation fingerprinting techniques to other domains, such as natural language processing and brain-computer interfaces.

Summary updated

BibTeX
@inproceedings{liu2025misclassification,
  author = {Weixing Liu and Shenghua Zhong},
  title = {{Misclassification-driven Fingerprinting for DNNs Using Frequency-aware GANs}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/843},
}
Figure