Radioactive Eye Information: Guarding Eye-Image Datasets through Radioactive Watermarking for Unauthorized-Use Detection

This paper explores radioactive watermarking as a technique for embedding invisible information in eye-tracking data, ensuring that any model trained on the modified samples retains an identifiable mark. Large-scale datasets have enabled robust deep learning models for appearance-based gaze estimation, but no reliable methods currently exist to detect unauthorized use of datasets. To address this, we evaluate radioactive watermarking, which embeds a watermark into eye data using pre-trained convolutional neural networks commonly used in gaze estimation models. We assess watermark robustness through gaze classification experiments, testing multiple neural architectures in different embedding and detection setups. Results demonstrate that training with watermarked data can be detected with high confidence, depending on the proportion of watermarked samples and the training setup. Detection is reliable with at least 10% watermarked data, while exceeding 15% degrades performance without significantly improving detection. Watermarks that retain high image quality preserve network performance and enable consistent detection.