Abstract
Deep learning models often perform multi-task classification (MTC), predicting multiple attributes from a single input. While efficient, this poses significant privacy risks when some attributes are sensitive. This paper presents the Multi-Trigger-Key (MTK) framework, a novel approach that utilizes a "lock and key" mechanism based on backdoor triggers to provide hierarchical access control. By distributing training and inference across four independent agents, MTK ensures that sensitive tasks remain protected unless an authorized trigger-key is applied to the input. Experimental results on the UTKFace dataset demonstrate that MTK effectively masks protected attributes while maintaining high accuracy for authorized users.
Introduction
Multi-Task Classification (MTC) is widely used in modern AI applications, such as facial analysis, where a single model can simultaneously predict gender, age, and race. However, the varying sensitivity of these attributes creates a privacy challenge. For instance, while gender might be considered public information, age or medical conditions are often highly sensitive. Existing privacy-preserving techniques, such as Differential Privacy, frequently suffer from accuracy trade-offs and lack granular, task-specific authorization mechanisms.
The Multi-Trigger-Key (MTK) framework addresses these limitations by introducing a hierarchical authorization system that persists throughout the model's lifecycle. It allows for "nested" access, where different users can be granted access to specific subsets of tasks based on their authorization level.
Methodology
The core of the MTK framework is the repurposing of backdoor triggers as a defensive security feature. Instead of using triggers to cause malicious behavior, MTK trains the model to produce random or useless outputs for protected tasks by default. These tasks are only "unlocked" when a specific pixel-pattern trigger-key is embedded into the input image.
System Architecture
The framework relies on four "semi-honest" agents to ensure distributed trust and prevent any single party from compromising the system:
- Data Distributor: Holds the raw training data.
- Secret Key Distributor: Manages the trigger-keys and handles user authorization requests.
- Assembler: Responsible for combining the raw data with the appropriate trigger-keys.
- Model Keeper: Trains and hosts the neural network.
Decoupling and Training
To prevent the model from inferring protected attributes through statistical correlations (e.g., guessing race based on age), the authors implement a decoupling step. A tuning parameter ($\tau$) is used to relabel highly correlated samples during training, breaking the statistical links that could lead to information leakage. This ensures that the model keeper, who never sees "clean" sensitive labels, cannot exploit task dependencies.
Experimental Setup
The framework was evaluated using the UTKFace dataset, containing over 20,000 images annotated with age, gender, and race. In the experiments, gender was treated as an unprotected task, while age and race were designated as secured tasks requiring specific triggers (S1 and C2).
Results
Experiments conducted on VGG16 and ResNet18 architectures confirmed the effectiveness of the MTK framework:
- Effective Masking: In the absence of the correct trigger-key, the model's accuracy on protected tasks dropped to near-random levels. For example, age classification accuracy fell from 67.9% to approximately 24%, and race classification fell from 81.9% to roughly 18%.
- Full Recovery: When the authorized trigger-key was applied, the model recovered its full predictive performance, matching the baseline of a standard, non-private model.
- Robustness and Sensitivity: The system proved robust across various trigger sizes (3x3 to 11x11 pixels) and color intensities. Notably, the model was highly sensitive to the exact location of the trigger, providing an additional layer of security against brute-force guessing.
Conclusion
The Multi-Trigger-Key framework offers a robust solution for multi-task privacy preservation by transforming backdoor vulnerabilities into a sophisticated authorization tool. By distributing knowledge across multiple agents and breaking task correlations, MTK provides strong privacy guarantees without sacrificing utility. Future research will focus on extending this framework to other domains such as video and natural language processing, as well as exploring more complex adversarial scenarios involving agent collusion.
Summary updated