TL;DR
This paper introduces a novel "box-free" method for verifying ownership of Generative Adversarial Networks (GANs). Instead of requiring specific input queries, it leverages the discriminator's learned representation of the generator's data distribution to create a unique "hypersphere" signature, making it robust against common attacks and applicable even when input selection is impossible.
What Problem Does This Paper Address?
Generative Adversarial Networks (GANs) are powerful tools for creating realistic images, but their valuable generators are susceptible to theft and illegal distribution. Existing methods for verifying ownership typically rely on a "black-box" setting, where an owner can query a suspicious model with specific inputs to trigger a hidden watermark or fingerprint. However, this approach has limitations, especially for generative tasks where input selection is not always feasible (e.g., unconditional image synthesis). Furthermore, these black-box methods are vulnerable to sophisticated attacks like input transformation and adversarial training, which can effectively remove or obscure the hidden functionality. The core problem this paper addresses is the need for a robust, input-agnostic method to verify GAN ownership, even when the adversary can freely query the model.
Methods
The paper proposes a "box-free" ownership verification scheme that capitalizes on the discriminator's role in GANs. The key idea is that a well-trained discriminator implicitly learns to represent the unique data distribution generated by its paired generator. The authors propose to:
- Empower the Discriminator: Instead of discarding the discriminator after GAN training, they leverage its learned representations. They introduce a "compactness loss" that encourages the discriminator to learn a hypersphere in its feature space that tightly encloses the distribution of generated data. This hypersphere acts as a unique signature of the generator.
- Preserve Discriminator Representations: GAN training can sometimes lead to the discriminator converging to a state where it loses its ability to extract meaningful representations. To combat this, they introduce a "Pearson Correlation Loss" during GAN training. This loss encourages the discriminator to implicitly reconstruct the latent space representation of generated images, thereby preserving its useful representational capacity.
- Box-Free Verification: For verification, a batch of images generated by a suspicious GAN is fed through the trained discriminator. The distance of these generated images' representations to the learned hypersphere is measured. A score is calculated based on this proximity. If the score deviates significantly from what's expected from the original model, ownership is claimed. This process requires no specific input selection, only the generated outputs.
Datasets
The authors evaluated their method on two popular datasets:
- LSUN: Used for image synthesis.
- CelebA: Used for both image synthesis and image-to-image translation.
They also used subsets of these datasets to investigate the impact of training data variations.
Key Insights and Ideas
- Discriminator as a Signature: The discriminator, often seen as a disposable component after GAN training, holds valuable information about the generator's learned distribution. This insight is central to the box-free approach.
- Hypersphere for Distribution Capture: Using a hypersphere to enclose the generator's data distribution in the discriminator's feature space provides a more robust and compact signature than traditional hyperplane-based methods.
- Pearson Correlation Loss for Stability: The Pearson Correlation Loss is a novel way to ensure the discriminator retains useful representational power during GAN training, preventing it from collapsing into a trivial state. This is crucial for the subsequent hypersphere learning.
- Box-Free is the Future: The paper strongly advocates for box-free verification, highlighting its inherent robustness against input-based attacks and its applicability in scenarios where input control is impossible.
Key Results
- High Verification Accuracy: The proposed method achieves high Area Under Curve (AUC) scores (often > 70%, and > 80% for SOTA architectures) in distinguishing between genuine and pirated GANs, even when only minor differences (like random seeds) exist.
- Robustness to Attacks: The method is shown to be immune to popular input transformation-based removal attacks and robust against model pruning. For instance, it can still verify ownership even when up to 40% of the model is pruned.
- Scalability: The approach generalizes well to various SOTA GAN architectures, with performance improving on more complex models.
- Efficiency: The verification process is described as time and memory-efficient, requiring only a single forward pass.
Limitations and Future Work
- Discriminator Disclosure: A key limitation acknowledged is that if the discriminator itself is compromised or disclosed, an adversary could potentially train a similar classifier or perform adversarial attacks. The authors note that in typical MLaaS (Model-as-a-Service) scenarios, the discriminator is not exposed.
- Extension to Other Models: The authors suggest future work could extend this distribution-capturing insight to other generative models, such as diffusion models.
- Evading the Proposed Method: They also acknowledge that exploring effective methods to evade their box-free verification scheme would be a valuable area for future research.
Summary updated