Biao Shen ^1* Su Yan ² Meilingzhi Chen ² Wei Kong ²

• ¹ Jiangsu Xinshun Energy Industry Group, Nanjing, China, Nanjing, China
• ² Nanjing Electricity Supply Industry General Corp, Nanjing, China

The increasing demand for privacy-preserving data processing techniques has led to significant advancements across various fields, including Named Entity Recognition (NER) within multimodal learning frameworks. This paper presents a novel approach to NER that ensures privacy by reasoning on noisy deep features across diverse data modalities, such as text, images, and audio. By introducing controlled noise into the data representation, our method preserves privacy while maintaining the effectiveness of NER across multi-modal inputs. The approach leverages advanced deep learning models, noise-adaptive attention mechanisms, and denoising autoencoders to accurately identify named entities in text and other modalities, despite the presence of noise. We evaluate our method on benchmark datasets, including CoNLL-2003, OntoNotes 5.0, and BioNLP-2004, demonstrating that it achieves competitive performance with state-of-the-art models in both unimodal and multi-modal settings. Additionally, our approach provides strong privacy guarantees, as evidenced by differential privacy metrics and resistance to data reconstruction attacks across multiple data types. Experimental results show that our method strikes a balance between privacy preservation and NER performance, offering a robust and generalizable solution for privacy-sensitive multi-modal NLP applications.