arXiv:2511.18829v1 Announce Type: new 
Abstract: Heart rate estimation from photoplethysmography (PPG) signals generated by wearable devices such as smartwatches and fitness trackers has significant implications for the health and well-being of individuals. Although prior work has demonstrated deep learning models with strong performance in the heart rate estimation task, in order to deploy these models on wearable devices, these models must also adhere to strict memory and latency constraints. In this work, we explore and characterize how large pre-trained PPG models may be distilled to smaller models appropriate for real-time inference on the edge. We evaluate four distillation strategies through comprehensive sweeps of teacher and student model capacities: (1) hard distillation, (2) soft distillation, (3) decoupled knowledge distillation (DKD), and (4) feature distillation. We present a characterization of the resulting scaling laws describing the relationship between model size and performance. This early investigation lays the groundwork for practical and predictable methods for building edge-deployable models for physiological sensing.

استكشف الباحثون تقطير نماذج التصوير الضوئي (PPG) الكبيرة المدربة مسبقًا إلى نماذج أصغر مناسبة لتقدير معدل ضربات القلب في الوقت الفعلي على الأجهزة القابلة للارتداء مثل الساعات الذكية وأجهزة تتبع اللياقة البدنية. تقيم الدراسة أربع استراتيجيات للتقطير: التقطير الصارم، التقطير اللين، التقطير المعرفي المنفصل، وتقطير الميزات، بهدف تلبية قيود الذاكرة والكمون لأجهزة الحافة.

Investigadores han explorado la destilación de grandes modelos preentrenados de fotopletismografía (PPG) en modelos más pequeños adecuados para la estimación en tiempo real de la frecuencia cardíaca en dispositivos portátiles como relojes inteligentes y rastreadores de fitness. El estudio evalúa cuatro estrategias de destilación: destilación dura, destilación suave, destilación de conocimiento desacoplada y destilación de características, con el objetivo de cumplir con las restricciones de memoria y latencia de los dispositivos de borde.

Des chercheurs ont exploré la distillation de grands modèles pré-entraînés de photopléthysmographie (PPG) en modèles plus petits adaptés à l'estimation en temps réel de la fréquence cardiaque sur des dispositifs portables tels que les montres intelligentes et les trackers de fitness. L'étude évalue quatre stratégies de distillation : distillation dure, distillation douce, distillation de connaissances découplée et distillation de caractéristiques, visant à respecter les contraintes de mémoire et de latence des dispositifs en périphérie.

Researchers have explored the distillation of large pre-trained photoplethysmography (PPG) models into smaller models suitable for real-time heart rate estimation on wearable devices like smartwatches and fitness trackers. The study evaluates four distillation strategies: hard distillation, soft distillation, decoupled knowledge distillation, and feature distillation, aiming to meet the memory and latency constraints of edge devices.

Towards Characterizing Knowledge Distillation of PPG Heart Rate Estimation Models

arXiv:2511.20456v1 Announce Type: new 
Abstract: Machine learning has become integral to Channel State Information (CSI)-based human sensing systems and is expected to power applications such as device-free activity recognition and identity detection in future cellular and Wi-Fi generations. However, these systems rely on models whose decisions can be subtly perturbed, raising concerns for security and reliability in ubiquitous sensing. Quantifying and understanding the robustness of such models, defined as their ability to maintain accurate predictions under adversarial perturbations, is therefore critical before wireless sensing can be safely deployed in real-world environments.
  This work presents a systematic evaluation of the robustness of CSI deep learning models under diverse threat models (white-box, black-box/transfer, and universal perturbations) and varying degrees of attack realism. We establish a framework to compare compact temporal autoencoder models with larger deep architectures across three public datasets, quantifying how model scale, training regime, and physical constraints influence robustness. Our experiments show that smaller models, while efficient and equally performant on clean data, are markedly less robust. We further confirm that physically realizable signal-space perturbations, designed to be feasible in real wireless channels, significantly reduce attack success compared to unconstrained feature-space attacks. Adversarial training mitigates these vulnerabilities, improving mean robust accuracy with only moderate degradation in clean performance across both model classes. As wireless sensing advances towards reliable, cross-domain operation, these findings provide quantitative baselines for robustness estimation and inform design principles for secure and trustworthy human-centered sensing systems.

تم إجراء تقييم منهجي لصلابة نماذج التعلم العميق ضد الهجمات العدائية، مع التركيز على أنظمة استشعار البشر المعتمدة على معلومات حالة القناة (CSI). تسلط هذه الدراسة الضوء على الحاجة الملحة لتحديد صلابة النموذج لضمان توقعات دقيقة في التطبيقات الواقعية، مثل التعرف على الأنشطة بدون جهاز والكشف عن الهوية.

Se ha llevado a cabo una evaluación sistemática de la robustez de los modelos de aprendizaje profundo frente a ataques adversariales, centrándose en los sistemas de detección humana basados en la Información de Estado del Canal (CSI). Esta investigación destaca la necesidad crítica de cuantificar la robustez del modelo para garantizar predicciones precisas en aplicaciones del mundo real, como el reconocimiento de actividades sin dispositivo y la detección de identidad.

Une évaluation systématique de la robustesse des modèles d'apprentissage profond face aux attaques adversariales a été réalisée, en se concentrant sur les systèmes de détection humaine basés sur l'information d'état de canal (CSI). Cette recherche met en évidence la nécessité cruciale de quantifier la robustesse des modèles pour garantir des prédictions précises dans des applications réelles, telles que la reconnaissance d'activités sans appareil et la détection d'identité.

A systematic evaluation of deep learning model robustness to adversarial attacks has been conducted, focusing on Channel State Information (CSI)-based human sensing systems. This research highlights the critical need for quantifying model robustness to ensure accurate predictions in real-world applications, such as device-free activity recognition and identity detection.

Towards Trustworthy Wi-Fi Sensing: Systematic Evaluation of Deep Learning Model Robustness to Adversarial Attacks

arXiv:2511.19260v1 Announce Type: new 
Abstract: Hunger and satiety dynamics shape dietary behaviors and metabolic health, yet remain difficult to capture in everyday settings. We present a Nutrition Photoplethysmography Language Model (NPLM), integrating continuous photoplethysmography (PPG) from wearables with meal descriptions. NPLM projects PPG into embeddings interpretable by language models, enabling joint reasoning over physiology and meal context. Trained on 19,340 participants and 1.1 million meal-PPG pairs, the model improved daily caloric intake prediction by 11% over text-only baselines, with accuracy maintained when 80% of meal text was removed. In an independent validation study (n=140) with controlled dining and detailed meal information, the model replicated these findings. These results demonstrate the value of integrating physiological measurements from consumer wearables with meal information for noninvasive dietary monitoring at scale.

تم تطوير نموذج لغة جديد للضوء الضوئي الغذائي (NPLM) يدمج بيانات الضوء الضوئي المستمر (PPG) من الأجهزة القابلة للارتداء مع أوصاف الوجبات لتحسين مراقبة النظام الغذائي. تم تدريب النموذج على مجموعة بيانات كبيرة تضم 19,340 مشاركًا و1.1 مليون زوج من الوجبات وPPG، وحقق تحسينًا بنسبة 11% في توقع استهلاك السعرات الحرارية اليومية مقارنة بالطرق التقليدية المعتمدة فقط على النص.

Se ha desarrollado un nuevo Modelo de Lenguaje de Fotopletismografía Nutricional (NPLM), que integra datos de fotopletismografía (PPG) continuos de dispositivos portátiles con descripciones de comidas para mejorar el monitoreo dietético. El modelo, entrenado con un conjunto de datos de 19,340 participantes y 1.1 millones de pares de comida-PPG, logró una mejora del 11% en la predicción de la ingesta calórica diaria en comparación con los métodos tradicionales basados solo en texto.

Un nouveau modèle de langage en photopléthysmographie nutritionnelle (NPLM) a été développé, intégrant des données de photopléthysmographie (PPG) continues provenant de dispositifs portables avec des descriptions de repas pour améliorer le suivi diététique. Le modèle, formé sur un ensemble de données substantiel de 19 340 participants et 1,1 million de paires repas-PPG, a atteint une amélioration de 11 % dans la prédiction de l'apport calorique quotidien par rapport aux méthodes traditionnelles basées uniquement sur le texte.

A new Nutrition Photoplethysmography Language Model (NPLM) has been developed, integrating continuous photoplethysmography (PPG) data from wearables with meal descriptions to enhance dietary monitoring. The model, trained on a substantial dataset of 19,340 participants and 1.1 million meal-PPG pairs, achieved an 11% improvement in predicting daily caloric intake compared to traditional text-only methods.

Towards Characterizing Knowledge Distillation of PPG Heart Rate Estimation Models

Was this article worth reading? Share it

LangWatch

LeapLife

Octofy