Privacy-Preserving Emotion Detection: Evaluating the Trade-Off Between K-Anonymity and Model Performance

In the realm of artificial intelligence, the pursuit of enhanced model performance has often prioritized the exponential growth of training data, sometimes relegating concerns about data privacy. This approach has fostered a perception that data privacy and the achievement of high-performance AI models are inherently opposing goals, particularly as digital fingerprinting is increasingly presented as essential for personalized experiences. This study aims to challenge this notion by demonstrating that even straightforward anonymization preprocessing techniques do not substantially alter the performance of machine learning models, regardless of their initial capabilities, while simultaneously safeguarding user privacy. We trained four different machine learning models: linear regression, linear ridge regression, a neural network, and a BiLSTM network, and evaluated their performance in data sets with varying levels of k-anonymity, specifically comparing results from a K-index of 1 to 52. Our findings indicate that, while certain trade-offs may exist, they should not be considered significant enough to deter the integration of anonymization techniques in machine learning and AI research. This work advocates for the routine adoption of anonymization practices, supporting the premise that robust model performance and strong data privacy are not mutually exclusive objectives.