Learning Realistic Patterns from Visually Unrealistic Stimuli: Generalization and Data Anonymization

Good training data is a prerequisite to develop useful Machine Learning applications. However, in many domains existing sets cannot be shared due privacy regulations (e.g., from medical studies). This work investigates simple yet unconventional approach for anonymized synthesis enable third parties benefit such data. We explore the feasibility of learning implicitly visually unrealistic, task-relevant stimuli, which are synthesized by exciting neurons trained deep neural network. As such, neuronal excitation can used generate synthetic stimuli. The stimuli train new classification models. Furthermore, we extend this framework inhibit representations that associated with specific individuals. use sleep monitoring both an open and large closed clinical study, Electroencephalogram stage data, evaluate whether (1) end-users create successfully customized models, (2) identity participants study protected. Extensive comparative empirical investigation shows different algorithms on able generalize same task as original model. Architectural algorithmic similarity between models play important role performance. For similar architectures, performance close using Accuracy difference 0.56%-3.82%, Kappa coefficient 0.02-0.08). Further experiments show provide state-ofthe-art resilience against adversarial association membership inference attacks.