This paper proposes a multi-scale knowledge distillation method based on self-supervised adversarial learning, named SAMKD, to address the challenge of ineffective knowledge transfer caused by the substantial scale difference between teachers and students in offline knowledge distillation. SAMKD leverages self-supervised and adversarial learning to further exploit the potential of intermediate multi-scale features and logits at the network’s output. Firstly, a supervised network learning model for multi-angle geometrically transformed images is introduced. Secondly, a multi-branch auxiliary network is designed to extract multi-scale features from the backbone network and obtain image logits. Finally, the binary game concept from generative adversarial networks is utilized for multi-stage adversarial training, enabling comprehensive knowledge transfer through the distillation process. Experimental analyses conducted on three challenging public datasets, CIFAR-10, CIFAR-100, and Tiny-ImageNet, demonstrate the strong competitiveness of the proposed method against other state-of-the-art knowledge distillation methods.