Animals and humans adapt to environmental changes by encoding prior experiences. While classical conditioning paradigms like fear conditioning have been the main focus in the learning and memory field, other forms of associative learning are gaining attention. One example is higher-order conditioning, mediated learning, which allows non-reinforced neutral stimuli to induce attraction or aversion through associations with directly reinforced cues. In rodents, sensory preconditioning, and second-order conditioning tasks have been used to assess higher-order conditioning. In this study, we developed a protocol for studying second-order conditioning in mice and a pipeline to analyze their behavioral responses computationally. The protocol involves direct conditioning between an electric shock and a low-salience stimulus (e.g., a light), which is then paired with a novel low-salience stimulus (e.g., a tone). These associative phases are followed by two probe tests where the low-salience stimuli are separately presented to characterize different behavioral responses in their presence (tone or light) to assess mediated and direct learning, respectively. Our assessment analyzes the video recordings with software known as DeepLabCut, a machine-learning approach based on convolutional neural networks. The output is further processed using Python scripts based on the open-source tool DeepOF. Our preliminary results show that this reproducible method can identify different behavioral patterns in second-order conditioning responses in mice. This research emphasizes the significance of introducing new computational tools to assess complex cognitive processes in mice. Combining this protocol with genetic and molecular approaches holds promise for advancing our understanding of the biological mechanisms underlying second-order conditioning.