Privacy-Aware Reconstruction and Profiling Using Advanced Contextual Embeddings

The proposed research aims to develop a secure and confidential system for identifying mental health disorders in individuals, with a focus on preserving the privacy of mental health data. This is achieved by generating encoded mental health profiles through the fine-tuning of a large language model (LLM) and applying advanced machine learning techniques. A transformer-based architecture within the LLM creates an embedding that serves as an encrypted key, incorporating Classification tokens. To maintain contextual accuracy, an autoencoder is trained to reconstruct these Classification tokens using the weights of a fine-tuned sentence Classification model. This design ensures that the network retains context while remaining secure. The system integrates demographic data and questionnaire responses into latent vectors that represent privacy-preserved mental health profiles. A fusion layer merges these vectors in the latent space to produce the final mental health profile. This privacy-conscious approach enables individuals to express mental health concerns securely, serving as a first step toward intervention. By employing natural language processing (NLP) techniques along with autoencoders, the research ensures data utility without compromising individual privacy. By leveraging this methodology, the research aims to improve mental health assessment while protecting sensitive data.