Edge intelligence is often assumed to improve privacy because raw sensor streams remain local and only constrained outputs (e.g., binary events, compressed embeddings, or coarse labels) are exposed. We show this assumption is misleading: even minimal on-device outputs can retain structured semantics that enable adversaries to infer sensitive behaviors under realistic context and access regimes. We introduce a game-based framework that separates two sources of leakage: statistical leakage, bounded by the information capacity of the observable channel, and algorithmic leakage, unlocked when adversaries exploit temporal coherence, multi-channel structure, or auxiliary context within that bound. Across embedded, smartphone, and multimodal sensing datasets, we find that a quantized motion interface can preserve 70–75% of behavioral structure (via information- and divergence-based scores) and enable ∼65% activity inference accuracy (about 4× random guessing) once temporal continuity is restored. Exposing richer continuous channels further increases in-domain leakage but becomes strongly placement- and device-specific, degrading transfer across sensors and datasets. Finally, we show that representation learning can induce cross-modal bridges that erode sensing-layer constraints, making seemingly low-sensitivity IMU signals more predictive of private semantic attributes associated with hidden high-fidelity modalities. Together, these results show that privacy in edge AI is shaped by an interaction between physical constraints and observable-interface design, motivating co-designed sensing, model, and evaluation choices that quantify and limit interface-induced leakage rather than assuming locality implies privacy.

Current sensor-gateway-cloud solutions for machine learning (ML) life cycle development require extensive coding, hardware, and software expertise. We present ST AIoT Craft, the first no-code/low-code online platform for training, deploying, and managing distributed artificial intelligence of things (AIoT) and big data. The framework provides user-interface dashboards and web applications for sensor data logging and management, dataset preprocessing and visualization, automatic tinyML for in-sensor processors and microcontrollers, and end-to-end Internet of things (IoT) system setup and monitoring. The tools allow developers to manage intelligent IoT devices, gateways, and cloud infrastructure directly from a web browser. ST AIoT Craft is accessible at https://staiotcraft.st.com/.

Inertial sensors provide a low-power and high-fidelity pathway for state estimation and sensor fusion. Inertial measurement units now feature on-chip processors for ultra-low-power information fusion, signal processing, and neural network-based classification at the extreme edge. However, accounting for domain shifts, personalized inference requirements, and application diversity makes adopting existing learning-enabled on-device training, classification, and fusion frameworks for on-sensor processors difficult. This paper introduces a method for personalized and on-device learning for on-chip classification, inference, and information fusion applications. The proposed framework automatically segments and stores quantized gravity vector image templates and axes variance information of motion artifacts during training. During inference, templates created from the time-series windows are matched against uniform blurred templates using the universal image quality index. An adaptive rep counting module robust to varying motion primitives counts repetitions of matched motion primitives. The framework requires no human-engineered parameters and allows for the personalization and addition of new motion artifacts. Our framework recognizes human activities with 96.7% test accuracy and achieves an average rep count error of 0.44, while reducing the memory usage by 1000-2000$\times$ over existing tiny machine learning on-device learning techniques, allowing on-sensor learning and inference under 8 KB of memory.

Precision agricultural robots require high-resolution navigation solutions. In this paper, we introduce a robust neural-inertial sequence learning approach to track such robots with ultra-intermittent GNSS updates. First, we propose an ultra-lightweight neural-Kalman filter that can track agricultural robots within 1.4 m (1.4–5.8× better than competing techniques), while tracking within 2.75 m with 20 mins of GPS outage. Second, we introduce a user-friendly video-processing toolbox to generate high-resolution (±5 cm) position data for fine-tuning pre-trained neural-inertial models in the field. Third, we introduce the first and largest (6.5 hours, 4.5 km, 3 phases) public neural-inertial navigation dataset for precision agricultural robots. The dataset, toolbox, and code are available at: https://github.com/nesl/agrobot.