"Adversarial Attacks on Encrypted Machine Learning Models"

Abstract

As machine learning models become increasingly integrated into sensitive domains, ensuring their security against adversarial attacks is paramount. Encrypted machine learning, which combines cryptographic techniques with model training, promises to safeguard data privacy during computation. However, recent studies reveal vulnerabilities where adversaries can manipulate encrypted inputs to induce erroneous model outputs without decryption. This abstract surveys existing adversarial attack methodologies tailored for encrypted machine learning models. It examines the efficacy of attacks exploiting various cryptographic protocols and model architectures. Additionally, it discusses mitigation strategies such as improved encryption schemes and adversarial training techniques to fortify models against these attacks. This exploration underscores the critical need for robust defenses in encrypted machine learning to uphold data confidentiality and model integrity in adversarial settings.