DistilBERT vs Gemma-3 270M for Mobile Cyberbullying Detection: A Comparative Study

Abstract

Cyberbullying poses critical urban safety challenges, with current cloud-based detection systems creating privacy concerns and infrastructure dependencies that particularly affect under-resourced communities. This paper presents a comparative study of DistilBERT and Gemma-3-270M for mobile cyberbullying detection, addressing sustainable on-device AI solutions for smart cities. Both models were evaluated on 95,000 samples from five hate-speech datasets and optimized using INT8 quantization for mobile deployment. DistilBERT achieved superior performance (86.01% accuracy, 86.32% F1-score) with minimal resource consumption (85 MB RAM, 64 ms latency). Gemma-3-270M demonstrated competitive results (78.48% accuracy, 79.23% F1-score) with generative capabilities for contextual explanations. Mobile deployment results confirm the feasibility of privacy-preserving, real-time detection on consumer devices. DistilBERT offers optimal efficiency for low-resource environments, while Gemma-3-270M provides broader versatility through text generation. This study was limited to English-language datasets and single-device evaluation, with future work planned to include multilingual datasets, cross-device validation, and formal statistical testing. Overall, this work supports equitable AI access in bandwidth-limited urban areas and advances sustainable smart-city development through efficient edge computing solutions.