隨著即時衛星影像分析需求的快速成長,將人工智慧(AI)功能直 接整合至低地球軌道(LEO)衛星平台,已成為重要的技術發展方向。 傳統衛星運作流程需將影像資料下傳至地面站進行分析,不僅產生延 遲,也增加通信成本,進而影響決策的時效性。為解決這些問題。本 文提出一種安全且輕量的可信 AI 推論架構,專門設計於資源受限且無 法實體接觸的 LEO 衛星環境中。並且整合了多項關鍵技術,包括利用 以 OPTEE 為基礎的韌體 TPM(optee ftpm)建立信任根,並透過 TPM 2.0 的平台組態註冊(PCR)測量機制與隨機數質詢協定進行遠端驗證, 以確保平台完整性與可信度。模型的安全傳輸則仰賴結合 TPM HMAC 的 AES 加密技術,防止在傳輸過程中遭到竄改或竊取。此外,推論過 程採用分區式設計,將敏感的模型層保護於可信執行環境(TEE)中, 而將其餘運算卸載至具備 GPU 加速能力的非安全區域(REE)執行, 兼顧性能與安全。實驗結果顯示,該系統可在合理時間內完成完整性 報告生成與模型驗證,同時透過 GPU 加速將推論延遲最高改善至 2.36 倍。在確保資料隱私與系統可信性的前提下,本文所提出的混合式架 構成功實現了 AI 模型從傳輸到執行的全程保護與可驗證性,為未來具 即時決策能力的衛星星座奠定了實用基礎。;With the growing demand for real-time satellite image analytics, inte- grating artificial intelligence (AI) capabilities directly into Low Earth Or- bit (LEO) platforms has become a critical technological frontier. Traditional satellite workflows rely on downlinking imagery to ground stations for analy- sis, incurring latency and communication overhead that hinder timely decision- making. This paper presents a secure and lightweight architecture for trusted AI inference on LEO satellites, addressing the unique constraints of resource- limited and physically inaccessible environments. Our system combines multiple core technologies: firmware TPM for es- tablishing a root of trust, Remote Attestation using TPM 2.0 PCR measure- ments and nonce-challenge protocols, AES encryption with TPM-HMAC for secure model delivery, and a partitioned inference framework leverag- ing GPU-accelerated REE execution and TEE-based protection for sensitive model layers. This hybrid architecture achieves end-to-end model protection and verifiable execution, laying a practical foundation for secure, real-time AI in future satellite constellations.