Privacy-Preserving Active Learning for sustainable aquaculture monitoring systems for extreme data sparsity scenarios

Описание и применение

Пайплайн, сочетающий активное обучение с регуляризацией разнообразия, DP‑SGD и агентскую оркестрацию, позволяет сократить количество необходимых меток примерно на 35 % при сохранении точности и обеспечении дифференциальной приватности в условиях экстремальной редкости данных и ограниченных ресурсов на борту датчиков.

Задача

Only a handful of labeled sensor readings were available from 40 deployed ESP32 water‑quality sensors, while the rest were miscalibrated, destroyed, or unlabeled, creating extreme data sparsity and privacy‑sensitive telemetry.

Результат

The pipeline reduced required labels by ~35% while maintaining F1 performance and providing formal differential privacy guarantees in an extreme sparsity, edge‑only setting.

Рабочий процесс

• Deploy ESP32 sensors with local 1D convolutional encoder to generate 32‑dim embeddings
• Use uncertainty sampling with diversity regularization to select unlabeled points for labeling
• Apply DP‑SGD with calibrated Gaussian noise during training
• Federate model updates from edge devices while accounting for privacy budget
• Run an agentic scheduler that monitors drift, budget, and decides next actions

Доказательства

Out of 40 sensors deployed across six ponds, only 11 returned labeled, trustworthy readings after the first month.

Источник

With lambda_div=0.5, the model reached the same F1 score with roughly 35% fewer labels than pure entropy sampling.

Источник

DP‑SGD: clip per‑sample gradients, then add Gaussian noise calibrated to the clipping norm.

Источник