Automated artificial intelligence (AI) accelerator generation platforms significantly simplify hardware design but introduce new attack surfaces. Attackers can embed malicious code into these platforms to implant hardware Trojans (HTs), leading to N-to-1 misclassification attacks. To counter this threat, we propose a lightweight pre-deployment defense method—Robust Training against Malicious Platforms (RTMP)—which aims to protect a specified target class by iteratively recalibrating the network’s decision boundaries to neutralize the impact of sensitive kernel combinations. Experiments on a state‑of‑the‑art automated accelerator generation platform show that, for ResNet‑18 models trained on CIFAR‑10 and ImageNet and for a VGG‑11 model trained on ImageNet, RTMP reduces attackable kernel groups by more than 90%, with a Top-1 accuracy degradation ranging from 0.1% to 2.4%, while increasing the platform’s search time by up to \(160\times \) . Moreover, RTMP substantially reduces the attack-success rate (ASR) of classical gradient-based adversarial attacks, achieving up to 35% reduction under PGD and up to 65% reduction under C&W. The code for this research is avaliable at https://github.com/islab-shi/RTMP.git .