Privacy-Preserving Deep Learning Models For Alcoholism Diagnosis Through EEG Data Analysis Using Differential Privacy Mechanisms

Alcoholism diagnosis through electroencephalogram (EEG) data analysis offers a promising alternative to traditional methods by identifying specific brain activity patterns associated with alcohol dependency. While deep learning techniques have demonstrated high accuracy in classifying EEG signals, privacy concerns related to sensitive medical data remain prevalent. Ensuring the privacy of patient data is critical for building trust and enabling the adoption of these tools in real-world clinical settings. This study develops deep learning models with enhanced privacy guarantees by incorporating differential privacy mechanisms, including (ϵ, δ)-Differential Privacy and Gaussian Differential Privacy (GDP). We compare their efficacy in preserving data privacy while maintaining model utility. Experiments show that convolutional and long-short-term-memory models optimized with Adam excel in utility and stability. GDP outperforms (ϵ, δ)-DP by requiring less noise, while DP-Adam surpasses DP-SGD in privacy and utility, particularly for fast convergence. Larger datasets further enhance this balance, emphasizing the importance of effective privacy mechanisms and sufficient data. By balancing privacy and utility, this work con-contributes a novel approach to privacy-preserving AI for sensitive health applications, emphasizing scalable models that maintain diagnostic accuracy.