In recent years, the use of contactless fingerprint recognition has increased, as it reduces the requirement for a specialized fingerprint acquisition sensor. This system offers better hygiene and is user-friendly in the authentication process, but these systems are also vulnerable to Presentation Attacks (PAs), ranging from silicone casts to printed replicas. As the major requirement of a contactless fingerprint system is to operate in any lightweight system, we have explored the use of two lightweight Convolutional Neural Network (CNN) architectures, MobileNetV3-Small and FBNetV3-B, on more than 11,000 bona fide and attack presentations drawn from ISPFDv1 and COLFISPOOF datasets, respectively. The performance of each model is evaluated on binarized versions of preprocessed images under the baseline and four Leave One Out (LOO) protocols, to test the generalization ability of the models for unseen attacks. We have achieved 0.32% Attack Presentation Classification Error Rate (APCER) at a fixed Bona Fide Presentation Classification Error Rate (BPCER) of 1% and the Detection Equal Error Rate (D-EER) of 0.52% on a binarized dataset. Results show that these architectures are capable of maintaining a perfect balance between performance and complexity for the specialized task of Presentation Attack Detection (PAD) in contactless fingerprint systems.