Termeh Taheri

🔬 Research

My current research focuses on enabling machines to reason about sound using symbolic audio representations. I’m exploring to improve how AI interprets and explains audio-based information. The project aims to bridge the gap between sound perception and natural language reasoning, contributing to more explainable and intuitive systems in the space of machine listening.

My broader interests include multimodal deep learning, machine listening, and signal processing, particularly systems that combine audio and language to enable more intuitive, interpretable AI.

📄 Selected Publications

Taheri, M., & Omranpour, H. (2024). Breast cancer prediction by ensemble meta-feature space generator based on deep neural network. Biomedical Signal Processing and Control, 87, 105382.
This work proposes EMFSG-Net, an ensemble meta-feature space generator designed to improve breast cancer classification from ultrasound images. By creating a more robust feature space through regression-based ensemble methods, the model overcomes challenges like overfitting and data scarcity — common in medical imaging. Unlike traditional approaches relying on trial-and-error data augmentation, EMFSG-Net boosts accuracy and efficiency without it. Evaluated on the BUSI dataset, the model achieved 97.96% accuracy and a 96.2% F1-score, outperforming existing deep learning models.
Omranpour, H., Mohammadi Ledari, Z., & Taheri, M. (2023). Presentation of encryption method for RGB images based on an evolutionary algorithm using chaos functions and hash tables. Multimedia Tools and Applications, 82(6), 9343–9360.
This paper presents a novel RGB image encryption technique that integrates chaos theory, evolutionary algorithms, and cryptographic hash functions. The method leverages sensitivity to initial values and randomized transformations to maximize security and resistance to attacks. A 256-bit hash is used to generate one-time keys, ensuring high entropy and strong defense against differential and plaintext attacks. The proposed algorithm is lightweight and fast, relying on simple operations like XOR and addition while maintaining robust encryption standards.