SWAP metrics optimization in mobile face anti-spoofing systems using knowledge distillation
Year: 2025
Type: article
Abstract: Face anti‑spoofing (FAS) on mobile devices demands models that are not only accurate but also fast, lightweight, and energy‑efficient – encapsulated by SWAP metrics (Speed, Weight, Accuracy, Power consumption). This paper investigates how knowledge distillation can optimize these SWAP metrics for neural networks in FAS. Large, high‑performing teacher models are distilled into compact student models that retain high accuracy while drastically reducing model weight and improving inference speed. Latest researches have shown that distilled FAS models can achieve accuracy on par with state‑of‑the‑art networks but with significantly lower computational cost, making real‑time mobile deployment feasible. The paper presents practical formulas for the knowledge distillation loss, comparative evaluations of models on SWAP criteria. It is concluded that knowledge distillation produces lightweight FAS models that run efficiently on mobile platforms (e.g., achieving nearly 7× faster inference for a distilled model with less than 1 M parameters, at approximately 99% of the teacher’s accuracy) while consuming a fraction of the power. In this article we outline future research directions – including multi-modal distillation and adaptive architectures, that could further push SWAP metrics optimization in this field.
Authors: Ostap Stets, Ihor Konovalenko
Institution: Ternopil Ivan Pului National Technical University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4413423948