TLDR: A research paper introduces a two-phase framework to address privacy concerns in AI-driven eye-tracking for interactive learning environments. Phase 1 demonstrates how student identities and diagnoses can be accurately backtracked using eye-tracking data. Phase 2 proposes a privacy-preserving solution featuring a dual-layer identity protection system with dummy IDs and administrator-only access, combined with Federated Learning. This approach allows for accurate diagnostic classification (99.3% in Phase 1, 99.4% in Phase 2) while effectively preventing identity tracing and securing student data.
Eye-tracking technology is becoming an increasingly powerful tool in interactive learning environments, offering insights into how students concentrate and learn. By observing where a person looks, this technology can help understand cognitive functions and even assist in diagnosing neurodevelopmental disorders. However, this capability also brings significant privacy risks, as the sensitive information collected could potentially be traced back to individual students, revealing their identity or even medical diagnoses.
A recent research paper, titled “Privacy Preservation and Identity Tracing Prevention in AI-Driven Eye Tracking for Interactive Learning Environments,” proposes a new human-centered framework designed to tackle this challenge. The framework aims to prevent identity backtracking while still allowing educators to benefit from the pedagogical advantages of AI-powered eye tracking. It integrates real-time data anonymization, ethical design principles, and compliance with regulations like GDPR to build trust and transparency in educational AI.
The researchers, Abdul Rehman, Are Dæhlen, Ilona Heldal, and Jerry Chun-Wei Lin, first demonstrated the potential for backtracking student IDs and diagnoses. They explored four scenarios in Phase 1:
Phase 1: Understanding the Vulnerabilities
1. Predicting Disorder Diagnoses: The models were trained on eye-tracking data from easier game levels and tested on more challenging ones. Even with varying task difficulties, the models achieved a high accuracy of 99.3% in predicting diagnoses like Moderate Intellectual Disability or Developmental Delay. This shows that diagnostic patterns can be learned and generalized across different cognitive contexts.
2. Predicting Student IDs Across Game Levels: In this scenario, the goal was to identify individual students based on their unique gaze patterns. While the models showed some ability to do this, with a Random Forest model achieving 63% accuracy, performance varied significantly among students. Some students were easily identifiable, while others were not, indicating that gaze consistency and task variation impact identification.
3. Predicting Student IDs with Random Data Split: When the dataset was randomly split for training and testing, the models achieved near-perfect accuracy (99.7% with a Decision Tree) in identifying students. This highlights that raw gaze features can be extremely effective for identification when data comes from a consistent distribution, but also underscores the serious privacy risks if such data is not protected.
4. Detecting New Students (Out-of-Sample): This scenario focused on identifying students who were not part of the initial training data. Using clustering methods like K-Means and outlier detection, the framework successfully identified new students, assigning them new IDs when their gaze patterns didn’t match existing profiles. This is crucial for managing privacy in dynamic learning environments where new students constantly join.
Also Read:
- Collaborative AI for Education: Addressing Privacy and Personalization with Federated Foundation Models
- Understanding Student Engagement: How Biosensors and Multimodal Analytics are Shaping Online Education
Phase 2: Building a Privacy-Preserving Solution
To counter the backtracking risks identified in Phase 1, the paper introduces a two-stage privacy-preserving framework. This phase focuses on securing student data and preventing identity tracing:
1. Dual-Layer Protection: This mechanism obfuscates true student identifiers with temporary dummy IDs. These dummy IDs are time-sensitive and periodically updated. Only authorized administrators can reverse this mapping using a secure key, ensuring that true identities are revealed only with proper authorization.
2. Federated Learning (FL): The framework employs Federated Learning, a decentralized machine learning approach. Instead of sending raw student data to a central server, each student’s device trains a local model using their own data. Only the model updates (not the raw data) are sent to a central server, which then aggregates these updates to improve a global model. This approach allows for collaborative learning and diagnostic prediction across multiple schools without ever centralizing sensitive student information.
The Federated Learning model, using a Deep Neural Network, achieved an impressive overall accuracy of 99.40% in preventing backtracking and securing the identity management system. This demonstrates that it’s possible to maintain high diagnostic accuracy while rigorously protecting student privacy.
The research also emphasizes the ethical considerations involved, noting that the pilot study complied with European ethical guidelines. All legal guardians provided informed consent, and children’s assent was obtained. Data was anonymized, and care was taken to ensure participant comfort and minimize intrusiveness.
In conclusion, this paper presents a robust framework that successfully secures students’ data and prevents identity backtracking in AI-driven eye-tracking systems for education. It offers a pathway to leverage advanced AI technologies for learning enhancement without compromising individual privacy. For more details, you can refer to the full research paper here.
