The growth of mobile financial transactions presents new challenges for fraud detection, where traditional and ML methods often miss emerging patterns. While Large Language Models (LLMs) offer advanced language understanding, they are typically too resource-intensive for mobile deployment and raise privacy concerns due to cloud reliance. This paper proposes a lightweight, privacy-preserving approach by fine-tuning and quantizing compact LLMs for on-device fraud detection from textual data. Models were optimized using Open Neural Network Exchange (ONNX) conversion and quantization to ensure efficiency. The fine-tuned quantized Llama-160M-Chat-v1 (bnb4) achieved 99.47% accuracy with a 168MB footprint, while fine-tuned quantized Qwen1.5-0.5B-Chat (bnb4) reached 99.50% accuracy at 797MB. These results demonstrate that optimized LLMs can deliver accurate, real-time fraud detection on mobile devices without compromising user privacy.