liBERTa: Local Intelligence via Browser Extensions for Real-Time Applications

As an application and service platform, the World Wide Web spans from simple informational websites to rich social media and Software-as-a-Service (SaaS) clients. While innovative capabilities are increasingly provided by Deep Learning (DL) Artificial Intelligence (AI) architectures such as pre-trained trans-formers, so far Web applications and services have integrated them only via cloud-based implementations. Deep-Learning-as-a-Service (DLaaS) is establishing itself for professional and personal use, with prevalent business models including pay-per-use and monthly subscriptions. With growing concerns over data privacy, response latency, and service costs, executing DL inference directly within the user's browser appears as a com-pelling alternative to cloud-based solutions. This paper introduces local intelligence via Browser Extension for Real- Time applications (liBERTa), a modular browser extension-based architecture for real-time client-side DL inference. By operating entirely within the browser, liBERTa reduces reliance on external servers. Its modular design consists of independent layers for data extraction, model inference, and results presentation, granting flexibility and adaptability across different kinds of applications and services. Experimental results from a case study on website privacy policy classification demonstrate the feasibility of the approach, showing that lightweight transformer models can achieve competitive accuracy while maintaining inference times suitable for real-world use on commodity hardware.