Abstract
Deep learning models represent significant intellectual property (IP), yet their inherent generalization capabilities often allow unauthorized users to repurpose them for new domains. This paper introduces MAP (MAsk-Pruning), a framework designed to protect model IP by surgically pruning parameters responsible for generalization to unauthorized domains. MAP operates effectively in source-free and data-free settings, maintaining high accuracy for authorized users while significantly degrading performance on unauthorized targets.
Introduction
The primary challenge in model IP protection is illegal knowledge transfer. A model trained for a specific authorized domain often generalizes to similar unauthorized domains, allowing third parties to exploit the model's capabilities without permission.
Existing protection methods typically require access to both original source data and unauthorized target data—a condition rarely met in practice due to privacy or data availability issues. MAP addresses this by providing protection across three increasingly difficult scenarios: Source-Available (SA-MAP), Source-Free (SF-MAP), and Data-Free (DF-MAP).
Methodology: The MAP Framework
MAP is built on the Inverse Transfer Parameter Hypothesis, which posits that a dense neural network contains a sub-network capable of maintaining high performance on an authorized domain while failing on unauthorized ones.
Surgical Mask-Pruning
Rather than retraining the entire model, MAP learns a target-specific binary mask while freezing original weights. This binary mask acts as a surgical filter, zeroing out parameters that facilitate transfer to unauthorized domains. This approach is more stable than continuous weight adjustment, which often leads to "catastrophic forgetting" of the primary task.
Handling Data Scarcity
To operate without original source data, MAP employs two specialized modules:
- Source Generator (SF-MAP): Synthesizes "pseudo-source" samples to help the model retain its original performance during the pruning process.
- Diversity Generator (DF-MAP): In the absence of any data, this module generates a wide variety of synthetic "neighborhood domains." These domains act as proxies for potential unauthorized targets, allowing the framework to shrink the model's generalization boundary effectively.
Evaluation and Results
The framework was evaluated on standard benchmarks including Digits (MNIST, USPS, SVHN), Object Recognition (CIFAR10, STL10), and the large-scale VisDA-2017 dataset across various architectures (VGG, ResNet, Swin-Transformer).
Key Findings
- Superior Protection: MAP consistently outperformed baselines like NTL and CUTI. In source-free settings, MAP maintained a source accuracy drop of only 8.8%, whereas competing methods saw drops exceeding 68%.
- Effective Blinding: While preserving source accuracy, MAP successfully reduced target domain performance to near-random levels (e.g., from ~60% to ~10% in certain tasks).
- Efficiency: Pruning redundant parameters occasionally resulted in slight performance gains on the authorized domain.
- ST-D Metric: The authors proposed the Source & Target Drop (ST-D) metric to quantify the trade-off between authorized performance and unauthorized protection, with lower values indicating superior IP protection.
Conclusion
MAP represents a significant advancement in model IP protection by enabling "blinding" in practical, data-constrained environments. While the quality of synthetic data generation remains a challenge for the data-free variant, MAP's surgical pruning offers a robust path for controlling the deployment and usage of high-value deep learning assets. Future research will explore extending this framework to complex tasks like semantic segmentation and natural language processing.
Summary updated