This paper describes the joint participation of Inria Défense et Sécurité and Storyzy to the ASVspoof5 challenge. We participated in the closed conditions of the audio deepfake detection and of the spoofing-aware speaker verification tracks with the goal of evaluating the performance of countermeasures with a fixed set of training attacks. The proposed countermeasure system is the combination of three models with different architectures and training algorithms, including the exploration of a self-supervised learning pretraining approach. Specific data augmentation strategies are introduced to increase robustness to numerical transmission and generalization to unknown attacks. The submitted system achieves a minDCF of 0.297 for track1 and a min a-DCF of 0.295 for track2. It has a very small calibration error (actDCF of 0.298) despite the presence of unknown codecs and adversarial attacks within the evaluation corpus.