Disinformasi digital yang semakin marak melalui penyebaran konten deepfake telah menjadi tantangan serius di era transformasi digital, khususnya di Indonesia. Penelitian ini bertujuan untuk mengembangkan dan membandingkan kinerja dua metode deteksi deepfake berbasis deep learning, yaitu Multi-Task Cascaded Convolutional Network (MTCNN) dan MobileNetV2. Penelitian difokuskan pada deteksi manipulasi citra wajah dengan memanfaatkan kekuatan ekstraksi fitur dari MTCNN dan efisiensi arsitektur ringan MobileNetV2. Proses penelitian melibatkan tahapan pre-processing data menggunakan metode ekstraksi wajah, pelatihan model klasifikasi biner, serta evaluasi performa melalui confusion matrix. Hasil penelitian menunjukkan bahwa MTCNN menghasilkan performa klasifikasi yang lebih baik dibandingkan MobileNetV2, dengan akurasi sebesar 84% dan nilai True Positive serta True Negative yang lebih tinggi. Sebaliknya, MobileNetV2 hanya mampu mencapai akurasi 78,5%, disertai tingkat False Negative dan False Positive yang lebih besar. Temuan ini menunjukkan bahwa MTCNN memiliki keunggulan dalam mendeteksi konten deepfake secara lebih akurat berkat kemampuan ekstraksi fitur wajah yang lebih presisi. Di sisi lain, MobileNetV2 menawarkan kecepatan dan efisiensi pengolahan data, namun kurang optimal untuk mendeteksi manipulasi detail pada wajah. Berdasarkan hasil penelitian, disimpulkan bahwa MTCNN lebih direkomendasikan sebagai metode deteksi deepfake dalam upaya mitigasi risiko disinformasi digital di Indonesia, terutama pada skenario yang menuntut akurasi tinggi. Increasingly widespread digital disinformation through the spread of deepfake content has become a serious challenge in the era of digital transformation, especially in Indonesia. This research aims to develop and compare the performance of two deep learning-based deepfake detection methods, namely Multi-Task Cascaded Convolutional Net-work (MTCNN) and MobileNetV2. The research focused on facial im-age manipulation detection by utilizing the feature extraction power of MTCNN and the lightweight architectural efficiency of MobileNetV2. The research process involved data pre-processing using face extraction methods, binary classification model training, and performance evalua-tion through confusion matrix. The results showed that MTCNN pro-duced better classification performance than MobileNetV2, with 84% accuracy and higher True Positive and True Negative values. In con-trast, Mo-bileNetV2 was only able to achieve 78.5% accuracy, along with higher False Negative and False Positive rates. The findings show that MTCNN has the advantage of detecting deepfake content more accurately thanks to its more precise facial feature extraction capabili-ties. On the other hand, MobileNetV2 offers speed and efficiency in data processing, but is less optimum for detecting manipulation of faci-al details. Based on the results, it is concluded that MTCNN is more recommended as a deepfake detection method in an effort to mitigate the risk of digital disinformation in Indonesia, especially in scenarios that demand high accuracy