A Study on Explainable AI-Driven Confidential Computing for Secure AI Workloads in Untrusted Cloud Infrastructures

Abstract

Hidden inside modern cloud systems, AI tasks now run-on shared machines where trust becomes a real problem. Instead of relying on clear safeguards, many depend on secure chips like Intel SGX or AMD SEV - these lock code away but give almost no view inside. Because users cannot see what happens during processing, checking if models stay safe feels nearly impossible. Without that window, sneaky threats creep in: copies of models stolen quietly, private details slipping out, inputs twisted by attackers. In settings built on suspicion rather than trust, missing oversight opens too many doors best left closed. Facing tough issues, this research introduces a secure computing setup powered by transparent artificial intelligence tools, combining smart interpretation methods with protected processing zones to boost trust and safety when running AI tasks on questionable cloud systems. Instead of just locking data away, it uses explanation strategies - like SHAP, LIME, or spotlighting key model choices - to watch how algorithms act inside shielded environments, offering clear reasons behind their moves. With insights built right into the protection layer, odd behaviors show up fast, hidden leaks get flagged early, meddling tries are spotted quicker - all while spelling out in plain terms why something got blocked or allowed. A first version got built, then tested against actual AI tasks across well-known cloud systems. What came out of testing shows the new method spots odd behavior well, without slowing things down much. Trust and clarity take a strong step up here, quite unlike older secure computing methods when put side by side. This study fits into wider efforts aiming at clearer, safer artificial intelligence within cloud setups. For fields like health care, banking, or vital operations, it holds real-world value - places were keeping data private and understanding AI decisions matter just as much for rules, trust, and fairness alike