<![CDATA[The rapid development of deep learning based face manipulation techniques has produced synthetic images that are increasingly realistic and visually indistinguishable from authentic ones. The deepfake phenomenon poses serious challenges to digital information authenticity and cybersecurity. This research presents a Systematic Literature Review (SLR) of publications from the 2020–2025 period to map trends, methodological approaches, and key challenges in machine learning and deep learning based image deepfake detection. Through an analysis of 24 empirical studies, this review identifies a shift in research direction from conventional convolutional architectures toward hybrid and attention based approaches that emphasize efficiency, adaptivity, and cross domain generalization. Findings show that although recent models such as Vision Transformer and hybrid CNN–LSTM are capable of achieving high accuracy under controlled conditions, their performance remains limited when tested on new domains. Key challenges identified include limited generalization against new manipulation types, vulnerability to image distortion and compression, and low transparency in model decision-making. This study fills research gaps by providing a comprehensive methodological map of architectural evolution, feature representation strategies, and evaluation metrics. Theoretically, this research expands the understanding of deepfake detection research dynamics, while practically, the results provide direction for developing adaptive, transparent, and efficient detection systems for real-time implementation.]]>
