Abstract
It is crucial to protect the intellectual property rights of DNN models prior to their deployment. The DNN should perform two main tasks: its primary task and watermarking task. This paper proposes a lightweight, reliable, and secure DNN watermarking that attempts to establish strong ties between these two tasks. The samples triggering the watermarking task are generated using image Mixup either from training or testing samples. This means that there is an infinity of triggers not limited to the samples used to embed the watermark in the model at training. The extensive experiments on image classification models for different datasets as well as exposing them to a variety of attacks, show that the proposed watermarking provides protection with an adequate level of security and robustness.
Introduction
Training state-of-the-art Deep Neural Networks (DNNs) is a resource-intensive process, requiring massive datasets and significant computational power. This investment makes the resulting models valuable intellectual property (IP). To protect this IP, researchers employ watermarking techniques. Traditional "black-box" methods often rely on backdooring—training a model to output a specific, incorrect label when presented with a fixed "trigger" image. However, these fixed-point triggers are vulnerable: if an attacker discovers them, they can fine-tune the model to "unlearn" the watermark.
"Mixer" addresses this vulnerability by moving from fixed trigger points to a secret manifold. By leveraging the Image Mixup technique, Mixer creates an asymmetric system where the samples used to verify ownership are never the same as those used during training, making the watermark exceptionally difficult for attackers to detect or remove.
Methods
The core innovation of Mixer is the application of Image Mixup to define a secret trigger space. While standard Mixup is a data augmentation technique used for generalization, Mixer repurposes it for security:
- The Secret Key: The owner defines a secret weighting vector $\lambda$ (the mixing proportions for each class) and a target output vector $\mu$ (the desired "surprising" label distribution).
- Embedding: During training, the owner injects "Mixer" samples created by blending one random image from every class according to $\lambda$. The model is trained to associate this specific blend with the secret label $\mu$.
- Verification: To prove ownership, the verifier uses the secret "recipe" ($\lambda, \mu$) to generate entirely new mixed images from a test set.
- Asymmetry: The specific images used to train the watermark are never used during verification. The model learns the underlying rule of the secret manifold rather than memorizing specific examples.
Experimental Results
The authors evaluated Mixer across three standard benchmarks: MNIST, CIFAR10, and a Transfer Learning (TL) scenario where a model pre-trained on ImageNet is fine-tuned for CIFAR10.
Robustness to Pruning
Mixer exhibits strong resilience against weight pruning. As shown in the figure below, the watermark recovery rate ($Rec_{tr}$ and $Rec_{ts}$) remains stable even as a significant portion of the network weights are removed. For MNIST and CIFAR10, the watermark is highly robust up to a pruning rate of approximately 0.45, after which the primary task accuracy (TA) also begins to degrade.
Resilience Against Model Attacks
The system was also tested against fine-tuning, quantization, and JPEG compression. As detailed in the following table, the watermark survives most standard post-processing steps. Notably, in the Transfer Learning scenario, the watermark recovery rate remained high (~88%), proving the watermark is deeply embedded in the model's logic. Adding the watermark did not significantly reduce the Test Accuracy (TA) on the primary task; for instance, on MNIST, accuracy remained at approximately 99.3%.
Conclusion
Mixer represents a significant step forward in DNN watermarking by shifting from static triggers to an infinite manifold of mixed images. This approach ensures that the watermark becomes a fundamental part of the model's decision logic, making it robust against pruning and fine-tuning. Future work will focus on mitigating "Usurper" threats—where an attacker might guess the secret parameters—and investigating resilience against active "overwriting" attacks.
Summary updated