Abstract
Training high-quality Generative Adversarial Networks (GANs) is a resource-intensive process, making these models valuable intellectual property (IP) and targets for theft. Traditional "black-box" verification methods rely on hidden watermarks or backdoors triggered by specific inputs. However, these methods are often fragile and easily bypassed by adversaries. This paper proposes a "box-free" ownership verification scheme that fingerprints a generator's unique output distribution using its paired discriminator. By repurposing the discriminator to map generated images into a learned hypersphere in feature space, the system can identify stolen models without requiring specific input triggers. Experimental results across multiple datasets and architectures demonstrate that this method is highly effective and robust against common model-removal attacks.
Introduction
The development of state-of-the-art GANs, such as those used for high-resolution image synthesis or deepfake generation, requires massive datasets, specialized hardware, and significant human expertise. This high cost of production makes GANs prime targets for IP theft. Protecting these models is critical for the AI industry.
Current verification techniques typically operate in a black-box setting. A defender embeds a secret "watermark" or "backdoor" into the model, which is activated only by a specific, secret input. This approach has significant drawbacks for GANs. Since many generation tasks use randomized latent vectors as inputs, a defender cannot easily "trigger" the watermark if an adversary uses their own random inputs. Furthermore, these methods are susceptible to "removal attacks," such as model pruning or input transformations, which can strip away the hidden watermark while maintaining the model's primary utility.
Methodology: Box-free Ownership Verification
To address the limitations of black-box methods, we propose a "box-free" approach. Instead of relying on hidden triggers, this method treats the generator's inherent output distribution as a unique fingerprint. The key innovation lies in repurposing the discriminator, a component typically discarded after the training phase.
Preserving Discriminator Utility
In standard GAN training, the discriminator often converges to a state where it provides little useful feature information once the generator becomes optimal. To prevent this, we introduce a Pearson Correlation Loss. This loss function encourages the discriminator to implicitly reconstruct the generator's original latent representation ($z$) from the generated images. This ensures the discriminator maintains a rich internal representation of the generator's specific stylistic artifacts.
Learning the Ownership Boundary
Following training, the discriminator's internal layers are used as an encoder. We train a secondary network to identify the smallest possible hypersphere in the feature space that encloses the majority of the generator's outputs. This hypersphere serves as the "ownership boundary."
Verification Process
To verify a suspicious model, the defender passes a batch of its generated images through the learned encoder. If the images consistently fall within the predefined hypersphere—indicated by a low Area Under Curve (AUC) score when compared to the original distribution—the model is flagged as a stolen copy.
Experimental Evaluation
The proposed scheme was evaluated using two industry-standard datasets:
- LSUN (Bedroom): For evaluating entire image synthesis.
- CelebA: For evaluating both image synthesis and image-to-image translation (attribute editing).
The method was tested across more than 10 different GAN architectures, including DCGAN, StyleGAN2, StyleGAN3, and StarGAN, to ensure broad applicability.
Results and Discussion
The box-free verification scheme demonstrated high precision and robustness across various tasks and attack scenarios.
Effectiveness
The system successfully distinguished between pirated models (identical weights) and independently trained models (homogeneous) with high confidence. For identical models, AUC scores were near 50%, indicating they share the same distribution. In contrast, models with different architectures or those trained on different data subsets yielded significantly higher AUC scores, typically between 70% and 80%.
Robustness to Attacks
The method proved resilient to common model-removal attacks:
- Model Pruning: Verification remains effective even if an adversary prunes up to 40% of the model's weights. Beyond this threshold, the quality of the generated images degrades to the point where the model loses its commercial value.
- Image Transformations: The verification process holds up against standard image "attacks," including the addition of noise, blurring, JPEG compression, and cropping.
Conclusion
By utilizing the discriminator as a "witness" to the generator's evolution, we have developed a stealthy and robust verification method. Because it relies on the inherent distribution of outputs rather than a hidden trigger, it is significantly harder for an adversary to detect or remove without destroying the model's performance.
A primary limitation is the security of the discriminator itself, which effectively acts as a "private key." If leaked, an adversary could potentially train a surrogate model to evade verification. Future research will explore extending this distribution-capturing approach to other generative frameworks, such as Diffusion Models, and investigating scalability for large-scale foundation models.
Summary updated