Given the rapid expansion of video content, the demand for efficient and scalable video summarization techniques has become increasingly critical. This work introduces a Convolutional Attentive Adversarial Network (CAAN)-based innovative unsupervised method for key frame detection and video summarization. The suggested technique models long-range dependencies between frames via a self-attention mechanism and uses a Fully Convolutional Sequence Network (FCSN) to collect temporal and global video properties. Our model learns to automatically give relevance scores to frames using a generative adversarial network architecture, allowing the extraction of important frames without annotated data. This strategy seeks to outperform a number of supervised techniques in addition to competing with currently used unsupervised methods. In order to overcome issues with annotation scarcity and improve the generalizability of video summarizing models, this work aims to offer a reliable unsupervised key frame detection solution.