Abstract
Mobile cloud offloading is indispensable for inference tasks based on large-scale deep models. However, transmitting privacy-rich inference data to the cloud incurs concerns. This paper presents the design of a system called PriMask, in which the mobile device uses a secret small-scale neural network called MaskNet to mask the data before transmission. PriMask significantly weakens the cloud's capability to recover the data or extract certain private attributes. The MaskNet is cascadable in that the mobile can opt in to or out of its use seamlessly without any modifications to the cloud's inference service. Moreover, the mobiles use different MaskNets, such that the collusion between the cloud and some mobiles does not weaken the protection for other mobiles. We devise a split adversarial learning method to train a neural network that generates a new MaskNet quickly (within two seconds) at run time. We apply PriMask to three mobile sensing applications with diverse modalities and complexities, i.e., human activity recognition, urban environment crowdsensing, and driver behavior recognition. Results show PriMask's effectiveness in all three applications.
Introduction
Deep learning models have become the backbone of modern mobile applications, from voice assistants to health monitoring. However, the computational demands of these models often exceed the capabilities of mobile hardware, necessitating "cloud offloading." In this paradigm, a mobile device sends raw data to an Inference Service Provider (ISP). This transmission exposes sensitive information—such as a user’s identity, gender, or precise location—to an "honest-but-curious" ISP that might attempt to reconstruct the data or extract private attributes beyond the scope of the requested service.
Existing privacy-preserving techniques often fall short in mobile environments. Homomorphic encryption is too computationally expensive, and standard neural network masking usually requires the ISP to retrain its proprietary models. Furthermore, many masking schemes are vulnerable to collusion: if the ISP obtains the masking parameters from a single compromised user, the privacy of the entire user base is often jeopardized.
The PriMask Framework
PriMask introduces a novel architecture designed to provide robust privacy without sacrificing utility or requiring cloud-side modifications. The system is built on three core components:
1. MaskNet: The On-Device Shield
The MaskNet is a tiny, on-device neural network (typically a simple Multi-Layer Perceptron) that transforms raw input into a masked format before it leaves the device. A key feature of PriMask is that it is cascadable. The MaskNet acts as a prefix to the ISP’s existing model (InferNet). Because the masking is designed to be compatible with the InferNet's expected input features, the ISP does not need to modify or retrain its "black-box" models.
2. HyperNet: Scalable Diversity
To prevent collusion attacks, every user must have a unique MaskNet. However, training a unique model for every user is computationally infeasible. PriMask solves this using a HyperNet—a generative model located at a trusted Privacy Service Provider (PSP). The HyperNet takes a random seed and instantly generates unique parameters for a user's MaskNet. This ensures that even if an ISP obtains the parameters for one user's mask, those parameters cannot be used to "decrypt" the data of any other user.
3. Split Adversarial Learning (SAL)
The training of the HyperNet is conducted via a novel protocol called Split Adversarial Learning. In this setup, the PSP and ISP cooperate without sharing raw data. The ISP provides gradients from its InferNet to ensure the masked data remains useful for the primary task (Utility). Simultaneously, the PSP trains the HyperNet against an internal AttackNet that attempts to reconstruct the private attributes from the masked data. This adversarial process forces the HyperNet to generate masks that hide private information while preserving the features necessary for inference.
Experimental Evaluation
The researchers evaluated PriMask across four distinct modalities to demonstrate its generalizability: Human Activity Recognition (HAR), Urban Environment Crowdsensing (UEC), Driver Behavior Recognition (DBR), and the MNIST digit dataset.
Key Results
- Efficiency: PriMask is highly practical for mobile deployment. Generating a new MaskNet takes less than two seconds on a workstation, and on-device execution takes only tens of milliseconds.
- Privacy vs. Utility: The system maintains high accuracy for the primary task while significantly reducing privacy risks. In the HAR dataset, PriMask reduced the ISP's ability to identify users from 84% to 17%, with only a 5% drop in activity recognition accuracy. In the DBR task, the system maintained 93.4% accuracy (compared to 98.4% on raw data) while ensuring that reconstructed images only showed vague contours rather than identifiable driver features.
- Collusion Resilience: Tests showed that even if 800 users colluded with the ISP to leak their masking keys, the attack success rate on non-colluding users remained low and stable, whereas non-heterogeneous systems would have been completely compromised.
Discussion and Future Work
PriMask represents a significant step toward practical, privacy-preserving mobile cloud inference. By introducing the role of a Privacy Service Provider (PSP), the system creates a trust hierarchy where a large, reputable entity (like a mobile OS provider) can verify privacy for smaller third-party app developers.
However, some challenges remain. The model currently assumes the PSP is fully trusted. Future research could explore decentralized methods to remove this dependency. Additionally, while SAL protects data, there is an ongoing discussion in the research community regarding "model leakage" during split learning, where one party might attempt to infer the architecture or weights of the other's model.
Conclusion
PriMask demonstrates that it is possible to achieve high-utility cloud inference while maintaining strong, collusion-resilient privacy. Its cascadable nature makes it uniquely suited for real-world deployment with existing, proprietary cloud models, offering a scalable solution for the growing mobile-cloud ecosystem.
Summary updated