Assisted Learning and Imitation Privacy

Xun Xian, Xinran Wang, Jie Ding, Reza Ghanadan

Abstract: In an increasing number of AI scenarios, collaborations among different organizations or agents (e.g., human and robots, mobile units) are often essential to accomplish an organization-specific mission. However, to avoid leaking useful and possibly proprietary information, organizations typically enforce stringent security constraints on sharing modeling algorithms and data, which significantly limits collaborations. In this work, we introduce the Assisted Learning framework for organizations to assist each other in supervised learning tasks without revealing any organization's algorithm, data, or even task. An organization seeks assistance by broadcasting task-specific but nonsensitive statistics and incorporating others' feedback in one or more iterations to eventually improve its predictive performance. Theoretical and experimental studies, including real-world medical benchmarks, show that Assisted Learning can often achieve near-oracle learning performance as if data and training processes were centralized.

Introduction

In the era of big data, the most valuable information is often siloed across different organizations. For example, a hospital might have patient records, while a separate clinical lab has the corresponding genetic data. Combining these would create a powerful predictive model, but strict privacy laws (like HIPAA) and proprietary concerns prevent them from simply swapping datasets.

Existing solutions like Federated Learning (FL) often require a central coordinator and homogeneous models, while Secure Multi-party Computation (SMC) is computationally intensive. Assisted Learning (AL) is a decentralized framework that allows organizations to collaborate on machine learning tasks without ever sharing private data, proprietary models, or even the specific nature of their tasks.

Methodology

The core of Assisted Learning is a two-stage protocol designed to be model-agnostic:

  1. Stage I (Learning Protocol): This is an iterative process. A "learner" (Alice) sends a query—typically the residuals (the errors her current model is making)—to an "assistant" (Bob). Bob treats these residuals as labels and fits a private model using his own local features. He then sends the fitted values back to Alice. Alice incorporates this feedback to update her model and repeats the process.
  2. Stage II (Prediction Protocol): Once training is complete, when a new data point arrives, Alice queries Bob for his model's prediction on that specific instance and combines it with her own to produce the final output.

Because AL only requires the exchange of residuals and fitted values, Alice could use a simple linear regression while Bob uses a complex Neural Network. They do not need to know each other's internal algorithms.

Experimental Evaluation

The researchers evaluated the framework using synthetic datasets (Friedman1) and the MIMIC-III Benchmarks, a massive real-world clinical database. They focused on two primary tasks: predicting Length of Stay (LOS) and Phenotype Classification.

MIMIC-III data structure and the application of Assisted Learning to medical divisions. (a) shows data tables used for specific tasks; (b) illustrates how Lab1 receives assistance from other divisions.

In the MIMIC-III setup, data is naturally partitioned into tables like Lab, Output, Input, and Prescriptions. AL allows these divisions to assist one another's specific predictive tasks without merging their underlying databases.

Discussion and Key Insights

  • Residuals as a Privacy Shield: By sharing residuals instead of raw data, the learner only reveals what they don't know yet, rather than the data they actually have.
  • The "Oracle" Target: The paper mathematically proves that for linear models, this iterative process converges to "near-oracle" performance—predictive accuracy almost as high as if all data had been merged into a single, centralized database.
  • Market-Driven Learning: The framework envisions a "statistical learning market" where various modules (hospitals, banks, cloud services) can act as either service providers or users, seeking or offering assistance based on their own goals.
The vision of a statistical learning market where various service modules assist a user module.

Key Results

The experimental results were highly promising. On both synthetic and real-world medical data, Assisted Learning performance closely tracked the oracle performance. In many cases, the gap between AL and a centralized model was negligible. Furthermore, AL significantly outperformed "Stacked Regression," as its iterative nature allows models to better account for synergies between features held by different parties. In the MIMIC-III Length of Stay task, error rates converged to the oracle level in just a few iterations.

Limitations and Future Work

While powerful, the framework faces challenges regarding Imitation Privacy, where a malicious learner might try to "steal" an assistant's model through repeated queries. Additionally, the iterative nature of the protocol introduces communication overhead that may be a bottleneck in low-bandwidth environments. Future work will explore optimal information exchange for classification tasks and methods for peer selection in large-scale learning markets.

Summary updated

Abstract

In an increasing number of AI scenarios, collaborations among different organizations or agents (e.g., human and robots, mobile units) are often essential to accomplish an organization-specific mission. However, to avoid leaking useful and possibly proprietary information, organizations typically enforce stringent security constraints on sharing modeling algorithms and data, which significantly limits collaborations. In this work, we introduce the Assisted Learning framework for organizations to assist each other in supervised learning tasks without revealing any organization's algorithm, data, or even task. An organization seeks assistance by broadcasting task-specific but nonsensitive statistics and incorporating others' feedback in one or more iterations to eventually improve its predictive performance. Theoretical and experimental studies, including real-world medical benchmarks, show that Assisted Learning can often achieve near-oracle learning performance as if data and training processes were centralized.
BibTeX
@inproceedings{xian2020assisted,
  author = {Xun Xian and Xinran Wang and Jie Ding and Reza Ghanadan},
  title = {{Assisted Learning and Imitation Privacy}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2020},
  url = {https://openreview.net/forum?id=2qAmyDhKjkU},
}
Figure