Learning Realistic Patterns from Visually Unrealistic Stimuli: Generalization and Data Anonymization (Extended Abstract)

Konstantinos Nikolaidis, Stein Kristiansen, Thomas Plagemann, Vera Goebel, Knut Liestøl, Mohan S. Kankanhalli, Gunn Marit Traaen, Britt Øverland, Harriet Akre, Lars Aakerøy, Sigurd Steinshamn

Introduction

In sensitive domains such as medicine, data privacy regulations like GDPR often create significant barriers to research. Sharing datasets like sleep recordings or brain scans is frequently prohibited, hindering the development of specialized AI models. While Generative Adversarial Networks (GANs) offer a way to create synthetic data, their focus on visual realism can inadvertently leak private information. This paper investigates a provocative alternative: do synthetic training samples actually need to look "real" to be effective for machine learning?

Methodology: The Teacher-Student Framework

The authors propose a four-step framework designed to transfer knowledge from a sensitive dataset to a new model without exposing the original data.

  1. Teacher Training ($h_T$): A "Teacher" model is first trained on the original, sensitive dataset.
  2. Stimuli Generation ($D_S$): Instead of mimicking the data's appearance, the researchers use Activation Maximization (AM). A generator network ($G_{AM}$) is optimized to produce inputs that trigger specific neurons in the Teacher model. These "stimuli" appear as unrecognizable noise to humans but are mathematically optimized to represent the features the Teacher uses for classification.
  3. Student Training ($h_S$): A "Student" model is trained using only these synthetic stimuli and the labels assigned by the Teacher.
  4. Evaluation: The Student is then tested on real-world data to evaluate its generalization capabilities.
The Teacher-Student framework: Step 3 involves training the student on synthetic data, while Step 4 evaluates it on real test data.

To further bolster privacy, the framework incorporates Inhibitory Stimuli. This technique penalizes the generator if the resulting stimuli contain features that could identify specific individuals from the original dataset.

An example of a visually unrealistic stimulus generated via Activation Maximization, which the teacher model identifies as the digit '9'.

Datasets and Key Insights

The approach was validated across diverse domains, including the MNIST handwritten digit benchmark, Sleep Apnea (SA) datasets (ECG and Nasal Airflow signals), and EEG sleep stage classification.

Several key insights emerged from this research:

  • Implicit Learning: Neural networks do not require "perfect" visual representations to learn. They can extract necessary statistical patterns from stimuli that appear as random noise to the human eye.
  • Privacy through Unrealism: The visual unrecognizability of the stimuli serves as a natural defense against human-led or automated re-identification attempts.
  • Task-Specific Knowledge Transfer: By focusing on the features required for a specific classification task, the method ignores extraneous, potentially sensitive information, inherently aiding anonymization.

Results and Privacy Evaluation

The Student models demonstrated remarkable generalization. The performance gap between models trained on real data versus those trained on AM-generated stimuli was minimal, ranging from only 0.56% to 3.82% depending on the task.

Beyond performance, the method proved robust against privacy threats. It showed significant defense against "membership inference attacks" and achieved state-of-the-art results in de-identification tests. Specifically, improvements in Cohen’s Kappa (measuring the prevention of re-identification) ranged from 0.02 to 0.41.

Conclusion and Future Directions

While a small performance gap remains compared to training on raw data, this work opens new avenues for privacy-preserving machine learning. Future research will focus on domain adaptation—tailoring models to new environments like home-based sensors—and scaling the approach to high-dimensional data such as genomic sequences or complex medical imaging.

Summary updated

Abstract

Good training data is a prerequisite to develop useful Machine Learning applications. However, in many domains existing data sets cannot be shared due to privacy regulations (e.g., from medical studies). This work investigates a simple yet unconventional approach for anonymized data synthesis to enable third parties to benefit from such anonymized data. We explore the feasibility of learning implicitly from visually unrealistic, task-relevant stimuli, which are synthesized by exciting the neurons of a trained deep neural network. As such, neuronal excitation can be used to generate synthetic stimuli. The stimuli data is used to train new classification models. Furthermore, we extend this framework to inhibit representations that are associated with specific individuals. Extensive comparative empirical investigation shows that different algorithms trained on the stimuli are able to generalize successfully on the same task as the original model.
BibTeX
@inproceedings{nikolaidis2022learning,
  author = {Konstantinos Nikolaidis and Stein Kristiansen and Thomas Plagemann and Vera Goebel and Knut Liestøl and Mohan S. Kankanhalli and Gunn Marit Traaen and Britt Øverland and Harriet Akre and Lars Aakerøy and Sigurd Steinshamn},
  title = {{Learning Realistic Patterns from Visually Unrealistic Stimuli: Generalization and Data Anonymization (Extended Abstract)}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.24963/ijcai.2022/806},
}
Figure