In pursuit of detecting dark matter signals, the Large Hadron Collider (LHC) at CERN has conducted proton-proton collisions to probe for these elusive particles, whose existence has been supported by astronomical observations. Despite extensive efforts by the CMS and ATLAS experiments, the direct detection of dark matter signals remains elusive. The current approaches employed for analyzing dark matter signatures utilize the cut-and-count method based on conventional techniques. This study introduces an alternative method for exploring dark matter signatures by utilizing fine-tuning of pre-trained models, such as ResNet-50, on 2D histograms generated from a combination of signal + background samples and background-only samples. By utilizing various signal-to-background ratios as benchmarks, an accuracy of about 90% for a signal-to-background ratio of 0.008 is achieved. This approach not only offers a more refined search for dark matter signals but also presents an efficient and effective means of analysis using machine learning techniques.