This work presents an unsupervised method for detecting ring-oscillator hardware Trojans in AES-128 FPGA implementations using side-channel power traces. The evaluated Trojan inserts an 11-stage LUT-based ring oscillator activated by an internal AES subnode, resulting in subtle but measurable deviations in switching activity. A dataset of 40,000 traces—20,000 clean and 20,000 Trojan across GF and LUT S-Boxes was analyzed using a deep autoencoder trained solely on clean data. Trojan traces exhibit significantly higher reconstruction error, enabling threshold-based classification. Results achieve 84.3% accuracy with clear separation between clean and Trojan samples, demonstrating the effectiveness of unsupervised reconstruction error analysis for scalable Trojan detection.