Abstract
As deep learning models become critical intellectual property, protecting them from unauthorized fine-tuning, pruning, and tampering is essential. Traditional protection methods utilize perceptual hashing to create model "fingerprints," but these often rely solely on model parameters (weights). This weight-centric approach leads to high false-positive rates, where distinct architectures may produce similar hashes. This paper proposes a novel structure-aware perceptual hashing framework that converts model architectures into Directed Acyclic Graphs (DAGs) and employs Graph Neural Networks (GNNs) to extract structural features. Our method achieves a low 2.69% false alarm rate and over 98% accuracy in identifying pirated models, significantly outperforming parameter-only baselines.
Introduction: The Vulnerability of Parameter-Only Hashing
Existing neural network protection schemes primarily focus on model parameters to generate unique identifiers. However, this focus ignores the model's "skeleton"—its architecture. This oversight creates a significant security gap: two entirely different models might accidentally generate similar parameter-based hashes, leading to high false-positive rates in piracy detection. Furthermore, adversaries can exploit this vulnerability to perform backdoor attacks, modifying a malicious model so its hash matches a verified, legitimate model.
Proposed Methodology
The core innovation of this work is the integration of structural identity into the perceptual hash. The authors propose a multi-stage pipeline that processes both the architecture and the weights of a white-box model.
1. DAG Extraction and Structure Hashing
To handle diverse neural network architectures, the system converts them into Directed Acyclic Graphs (DAGs). This approach simplifies the model representation into a three-layer design (input, intermediary layers, and output nodes), reducing computational complexity from $O(mn)$ to $O(m+n)$. A Graph Isomorphism Network (GIN), pre-trained on molecular graph data and fine-tuned via contrastive learning, then extracts structural embeddings from these DAGs. This ensures that similar architectures result in similar embeddings while distinct structures are clearly differentiated.
2. Parameter Hashing and Fusion
Simultaneously, a parameter hashing module processes the model's weights. To ensure efficiency and reduce redundancy, the weights are pruned before being passed through a dual-branch network. Finally, a "neural fusion" merge layer combines the structural and parameter hashes into a single, robust 100-bit perceptual hash.
Experimental Evaluation
The framework was validated using a diverse dataset of 2,569 models, including popular architectures like ResNet, VGG, ShuffleNet, and MobileNetv2, trained on standard vision tasks (CIFAR-10/100, MNIST, FashionMNIST).
Key Results
- Enhanced Precision: The structure-aware approach reduced the False Alarm Rate (FAR) to 2.69%, a significant improvement over the 9.50% FAR observed in parameter-only baselines.
- High Detection Sensitivity: The system maintained over 98% accuracy in detecting pirated models that had undergone fine-tuning or pruning.
- Robustness: Even when subjected to significant parameter tampering (e.g., Gaussian noise), the recognition accuracy remained above 90%.
- Fusion Efficacy: Ablation studies confirmed that learned "neural fusion" for merging hashes outperforms simple concatenation, effectively capturing the complex relationship between a model's structure and its weights.
Conclusion and Future Directions
By treating a model's structure as a core component of its identity, this research provides a more reliable method for neural network authentication and piracy detection. While currently optimized for white-box scenarios, future work aims to extend this framework to Large Language Models (LLMs). Given the immense scale of LLMs, the authors propose analyzing them as distributional patterns in parameter space to manage their architectural complexity while maintaining structural awareness.
Summary updated