Abstract
PrivFusion introduces a decentralized architecture for fusing fine-tuned AI models without sharing raw weights or relying on a trusted third party. By treating neural networks as graphs and employing Local Differential Privacy (LDP) alongside federated graph matching, PrivFusion enables secure collaborative model improvement. Evaluation on MNIST, CIFAR-10, and real-world medical imaging datasets demonstrates that PrivFusion achieves high utility—nearly matching non-private state-of-the-art methods—while providing robust privacy guarantees and handling non-IID data effectively.
Introduction
In the era of "Model-as-a-Service" (MaaS), users often download pre-trained models and fine-tune them on private, local data. To create even better models, these users might want to "fuse" their individual updates. However, two major roadblocks exist: privacy risks and centralization. Sharing model weights directly can lead to "inversion attacks," where an adversary reconstructs sensitive training data. Furthermore, current fusion methods often require a trusted central server, creating a single point of failure and a "presumed trust" that many organizations, such as hospitals, cannot afford.
PrivFusion addresses these challenges by allowing multiple parties to fuse their fine-tuned AI models into a single, high-performing model without ever sharing raw weights or relying on a middleman. By combining graph matching with Local Differential Privacy (LDP), it protects sensitive training data while maintaining model utility, even in complex fields like collaborative healthcare.
Methods
PrivFusion treats neural networks as graphs where neurons are nodes and weights are edges. The decentralized process follows four main steps as illustrated in the workflow below:
- Model Graph Construction: Each user converts their local model into a graph representation.
- Local Model Perturbation (LDP): Before sharing, users apply Local Differential Privacy. They use a Laplace mechanism to add noise to neuron activations and a MultiBit mechanism to perturb the weights, ensuring the shared "noisy" graph does not leak specific data samples.
- Decentralized Federated Graph Matching: Users calculate the "affinity" (similarity) between their noisy graphs to find a permutation matrix. This map identifies corresponding neurons across models without a central coordinator.
- Private Model Fusion & PFA: Once aligned, the models are averaged. To combat performance drops from noise, a Perturbation Filter Adapter (PFA) is introduced. This includes a "Smoothing Filter Unit" (SFU) that uses a Gaussian Error Function to standardize results, helping the fused model regain accuracy.
Experimental Evaluation
The researchers evaluated PrivFusion on three distinct setups:
- MNIST: A standard handwritten digit benchmark using fully connected and CNN architectures.
- CIFAR-10: A complex image dataset used to test classic CNNs.
- Real-World Healthcare: A dataset of 10,217 color fundus images for diagnosing age-related macular degeneration, collected from 17 specialized hospitals. This tested performance on "non-IID" data where different hospitals have varying patient demographics.
Key Results
- High Fidelity: On MNIST, PrivFusion achieved 93.01% accuracy, nearly matching non-private state-of-the-art methods like GAMF (93.17%) and OTFusion (93.13%).
- Robustness to Data Imbalance: In the healthcare dataset, PrivFusion maintained an accuracy of ~80%, outperforming individual local models even when data was highly unbalanced.
- Controllable Privacy: By adjusting the privacy budget ($\epsilon$), users can precisely balance noise tolerance against the required level of security.
Discussion and Insights
Analysis revealed that perturbing activation values has a significantly larger impact on fusion quality than perturbing weights. This suggests that protecting "how the model thinks" (activations) is the most critical aspect of privacy-preserving fusion. PrivFusion’s hybrid approach—using different mechanisms for activations, weights, and fusion—allows it to achieve a superior privacy-utility trade-off compared to single-mechanism methods.
Unlike Federated Learning, which requires many rounds of communication and active training, PrivFusion merges already-trained models without needing the original training data or further gradient updates.
Limitations and Future Work
While efficient for modest networks, graph matching involves a $O(d^4)$ memory cost for quadratic assignment problems, making scaling to massive foundation models (like large LLMs) a challenge. Future work will focus on "noise-adaptive" perturbation techniques and extending this decentralized approach to large-scale collaborative scenarios involving hundreds of participants.
Summary updated