Masterarbeit
Validation of inference in Alaas systems
Completion
2025/05
Research Area
Students
Anas Abubader
Advisers
Abubaker Gaber
Dr.-Ing. Sebastian Heil
Description
In the era of AI-driven solutions, the reliance on artificial intelligence (AI) services has become ubiquitous across industries, ranging from healthcare to finance and beyond. These services are often marketed with promises of high performance, accuracy, and reliability, which form the foundation of user trust. However, ensuring that these promises align with actual service delivery is a persistent challenge. Misrepresentation of the service claimed to be provided or covert downgrades in quality can lead to financial loss, ethical dilemmas, and reduced confidence in AI as a transformative tool. Such discrepancies highlight the critical need for mechanisms that validate the commitments of AI providers.
The proposed middleware aims to address the critical issue of trust and accountability in AI service delivery by serving as a neutral validator of AI service claims. In environments where users lack direct access to verify the service claimed to be provided by AI providers, this middleware offers an innovative solution through continuous, transparent monitoring. By implementing zero-knowledge proof (zk-proof) techniques, the middleware can verify the AI provider’s compliance with the service claimed to be provided without accessing or compromising proprietary information. Through frequent, randomised validation checks, the middleware can identify any "cheating attempts"—instances where a provider delivers a lower-tier service than claimed—and promptly record them in a database. Over time, these records can be used to calculate a "trust percentage" for each provider, offering users a measurable indicator of reliability. This architecture not only supports transparency but also protects sensitive data by verifying the service claimed to be provided without revealing internal processes, thereby fostering trust between AI providers and users while upholding confidentiality.
