Cloud-based machine learning (ML) platforms provide scalable GPU resources but limit user oversight, exposing training processes to risks such as computation skipping and model poisoning. Ensuring training integrity is critical to counter these threats by verifying proper execution. However, current solutions often lack real-time responsiveness and cost-effectiveness, resulting in delayed threat detection and higher expenses that restrict their practical deployment. To overcome these challenges, we propose LightGuard, a secure, real-time, and cost-efficient verification framework tailored for cloud-based ML. LightGuard leverages Trusted Execution Environments (TEEs) on cloud platforms to perform periodic duplicated-execution checks, ensuring real-time integrity assurance. To address the performance bottlenecks of TEEs, LightGuard incorporates SmartFetch, an adaptive scheduling algorithm that dynamically optimizes checkpoint intervals and coordinates TEE and GPU resources. This enables asynchronous training and verification under strict security and latency constraints. Theoretical analysis confirms that LightGuard delivers strong security guarantees and maintains high performance. Experiments show that it incurs only 12% latency overhead compared to unprotected training, detects attacks 4–5 \(\times \) faster than prior methods, and prevents 40–50 compromised epochs under practical 5–20% verification budgets. Offering a formal security guarantee with error rate \(\le \) 1%, LightGuard significantly improves the trustworthiness of cloud-based ML systems.