XNN: Paradigm Shift in Mitigating Identity Leakage within Cloud-Enabled Deep Learning

Kaixin Liu, Huixin Xiong, Bingyu Duan, Zexuan Cheng, Xinyu Zhou, Wanqian Zhang, Xiangyu ZhangarXiv 2024 · DOI · arXiv

Abstract

As deep learning tasks are increasingly outsourced to untrusted cloud platforms, the risk of "identity leakage"—where sensitive biometric data can be linked back to individuals—has become a critical concern. This paper introduces XNN and XNN-d, two frameworks designed to secure the training and inference phases of face recognition models. By employing randomized feature obfuscation and adversarial distillation, these methods reduce attack success rates to near-zero levels while maintaining high recognition utility.

Introduction

In the contemporary landscape of "Cloud-Enabled Deep Learning," organizations frequently leverage third-party cloud infrastructure for computationally intensive model training and deployment. However, this outsourcing introduces significant privacy vulnerabilities. Even when raw images are visually obscured (e.g., through blurring), the underlying mathematical features extracted by deep neural networks often remain identifiable. An adversary with access to a "gallery" of known identities can perform feature matching to deanonymize users with high precision.

Existing privacy-preserving techniques often suffer from a steep utility-privacy tradeoff or remain susceptible to reconstruction attacks. XNN addresses these challenges by ensuring that identity-sensitive information is obfuscated before it ever reaches the untrusted cloud, while preserving the structural relationships necessary for accurate model performance.

Methodology

The proposed solution comprises two distinct pipelines tailored for different stages of the machine learning lifecycle.

XNN: Secure Training Phase

XNN focuses on the fine-tuning stage. Before data transmission, the data owner processes images through a feature extractor followed by an ObfNet. This network performs two primary operations: random patch permutation and random matrix multiplication.

The XNN architecture featuring ObfNet for feature obfuscation.

The parameters governing these transformations are kept secret by the data owner. The cloud provider then trains a recognition network (RecNet) on these scrambled features. Because the relative topology of the feature space is preserved, the model learns to distinguish between classes without accessing the true identity of the subjects.

XNN-d: Secure Inference Phase

XNN-d (where 'd' denotes distillation) is designed for the deployment phase. It utilizes a generator to inject adversarial noise into features, specifically engineered to thwart unauthorized identity recognition. To maintain the functionality of the authorized cloud server, a distillation technique is employed: a "student" recognition network is trained to interpret the noised features by learning from a "teacher" network that has access to clean data. This allows the server to perform its task while remaining "blind" to the actual identity of the user.

Experimental Evaluation

The frameworks were evaluated using large-scale face recognition benchmarks, including CelebA, CASIA-WebFace, MSRA, and VGGFace2, totaling over 3 million images. To simulate real-world threats, auxiliary datasets like FaceScrub and IMDb-Face served as the adversary's gallery for testing identity matching.

Key Results

The evaluation demonstrates that XNN significantly outperforms baseline methods and existing privacy frameworks like InstaHide and NeuraCrypt. On the CelebA dataset, XNN reduced the Attack Success Rate (ASR) from a vanilla baseline of 67.77% to a mere 0.17%. Crucially, this security gain came with minimal impact on utility; recognition accuracy only decreased from 67.77% to 64.96%.

Discussion and Insights

A central finding of this research is that visual data preservation is not a prerequisite for effective model training. The "topology preservation" principle allows models to learn discriminative boundaries even in a scrambled feature space.

t-SNE visualization showing preserved feature clusters after obfuscation.

As shown in the t-SNE embeddings, while obfuscation shifts the absolute positions of identity clusters, it maintains the relative distances between them. This allows the cloud-based model to distinguish "Person A" from "Person B" without ever "seeing" a recognizable representation of either.

Limitations and Future Directions

While XNN provides robust protection, it is an obfuscation-based method rather than a formal cryptographic proof. Its security relies on the secrecy of the transformation parameters. Furthermore, if an adversary obtains a sufficient number of original and obfuscated feature pairs, they might theoretically approximate an inverse function. Future research will focus on extending this paradigm to other sensitive domains, such as medical imaging and financial data analysis.

Conclusion

XNN and XNN-d represent a significant step forward in balancing privacy and utility in cloud-enabled deep learning. By moving the security boundary to the data owner's side and utilizing topology-preserving obfuscation, these frameworks effectively mitigate the risk of identity leakage in untrusted environments.

Summary updated

Abstract

In the domain of cloud-based deep learning, the imperative for external computational resources coexists with acute privacy concerns, particularly identity leakage. To address this challenge, we introduce XNN and XNN-d, pioneering methodologies that infuse neural network features with randomized perturbations, striking a harmonious balance between utility and privacy. XNN, designed for the training phase, ingeniously blends random permutation with matrix multiplication techniques to obfuscate feature maps, effectively shielding private data from potential breaches without compromising training integrity. Concurrently, XNN-d, devised for the inference phase, employs adversarial training to integrate generative adversarial noise. This technique effectively counters black-box access attacks aimed at identity extraction, while a distilled face recognition network adeptly processes the perturbed features, ensuring accurate identification. Our evaluation demonstrates XNN's effectiveness, significantly outperforming existing methods in reducing identity leakage while maintaining a high model accuracy.
BibTeX
@misc{liu2024xnn,
  author = {Kaixin Liu and Huixin Xiong and Bingyu Duan and Zexuan Cheng and Xinyu Zhou and Wanqian Zhang and Xiangyu Zhang},
  title = {{XNN: Paradigm Shift in Mitigating Identity Leakage within Cloud-Enabled Deep Learning}},
  year = {2024},
  eprint = {2408.04974},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2408.04974},
}
Figure