<p>Rapid advancement in deepfake technology poses serious threats to various aspects of society, including false information, privacy breaches, manipulation of content, etc. Today, there exist various methods based on machine learning, deep learning, statistical analysis, signal processing, etc., for deepfake image detection. Fine-grained manipulations like texture distortions or blending artifacts are often overlooked by existing methods. In this paper, we have introduced a model based on the Contractive Autoencoder for deepfake image detection. The Contractive Autoencoder can learn robust and invariant features through the contractive penalty in its loss function that ensures fine-grained manipulations. Unlike existing methods, the proposed method leverages the inherent ability of the Contractive Autoencoder to capture high-dimensional data structures and extract discriminative features, thus enhancing detection accuracy. We have applied the proposed method to the DFDC faces dataset (Approximately 20,000 images), which is online available on Kaggle. The model has achieved the highest accuracy of 97.47%. It has also been evaluated for the parameters precision, recall, f1-score, and ROC curve, and found good results.</p>