A Novel Polynomial Activation for Audio Classification Using Homomorphic Encryption

With the increasing popularity of audio recording devices and recognition technology, audio data is increasingly used in speech recognition, event detection, and biometric authentication. Audio data often contains sensitive information, raising privacy concerns when using artificial intelligence on cloud platforms. Homomorphic encryption (HE) addresses this problem by allowing direct computation on encrypted data without decryption. However, HE has high computational costs, especially for deep learning models, which require many nonlinear operations, such as activation functions. Traditional HE methods have difficulty handling these nonlinear operations, affecting performance and accuracy. To overcome this limitation, we propose optimized polynomial-degree activation functions that enhance the compatibility of HE with deep learning models while maintaining high performance. Experimental results on musical instrument and audioMNIST datasets confirm the effectiveness of our method, highlighting its promising potential for secure audio data processing in various applications.