ARCHIVES

A model, in artificial intelligence, is a computational system trained on data to recognize patterns, make predictions, or generate meaningful insights. Model-as-a-Service (MaaS) allows users to remotely access such pre-trained models through the cloud, enabling applications like image classification, financial forecasting, and intelligent decision-making without requiring local hardware or training infrastructure. However, this cloud-based architecture requires users to upload sensitive data to external servers, creating serious privacy risks. Such data may be exposed to inference leakage attacks, where adversaries can deduce information about user inputs or extract characteristics of the deployed model, posing substantial security risks. To address these concerns, this project presents a privacy-preserving MaaS framework built on Fully Homomorphic Encryption (FHE). In this architecture, user data is encrypted before being sent to the server, allowing the cloud to evaluate AI models directly over ciphertext. The server processes the encrypted query and returns an encrypted inference result, which only the user can decrypt locally. At no point does the service provider gain access to plaintext input or output, ensuring strong confidentiality for both user data and model intelligence. This encrypted-inference workflow is optimized to support real-time applications such as secure facial verification and rapid predictive assessments while maintaining robust privacy guarantees. By eliminating exposure of sensitive information and preventing model inversion attacks, the proposed FHE-enabled MaaS framework enhances trust, strengthens security, and enables privacy-centric deployment of AI models on the cloud.