Edge-Compatible Deep Learning for Biometric Identification from Wrist-Worn PPG Signals
30th International Conference on Information, Communication and Automation Technologies, ICAT 2025, Sarajevo, Bosna-Hersek, 2 - 04 Eylül 2025, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/icat66432.2025.11189098
- Basıldığı Şehir: Sarajevo
- Basıldığı Ülke: Bosna-Hersek
- Anahtar Kelimeler: biometric identification, deep learning, edge AI, EER, LSTM, PPG, TensorFlow Lite, Transformer
- Boğaziçi Üniversitesi Adresli: Evet
Özet
Wrist-worn photoplethysmography (PPG) offers an unobtrusive biometric modality, yet state-of-the-art models rarely address the memory and runtime constraints of embedded hardware. We conduct a systematic benchmark of CNN, LSTM, Transformer, and hybrid architectures on three public datasets, UBFC-Phys, ECSMP, and EMOGNITION, encompassing 166 unique subjects and over 26,000 PPG windows. Models are evaluated under a unified preprocessing pipeline and then converted to TensorFlow Lite to gauge deployability. A lightweight LSTM-Transformer architecture achieves 99.14% accuracy and a global Equal Error Rate of 0.19%. After TFLite conversion, accuracy drops by less than 0.1% while model size shrinks from 730 kB to under 100 kB. Inference latency on a representative resource-constrained CPU ranged between 13-50 s per 10s window using single-threaded TFLite; we therefore outline quantization and hardware-aware optimization as future work.