AI-generated image detection algorithm based on classical-quantum hybrid neural network

The advance in generative artificial intelligence (AI) has led to increasingly realistic image synthesis, which from a forensic perspective, also spawned the demand for AI-generated image detections. Achieving detection generalizability across various deep generative models is crucial as different deep generative models are emerging. The key to achieving generalizability is to design better architectures for feature extraction and representation. Noteworthy, quantum neural networks (QNN), due to their larger representation space and inherent parallelism, have already shown their superior performance in feature extraction. Therefore, in this paper, we introduce a delicately designed QNN and combine it with the classical Swin Transformer V2, proposing an AI-generated image detection algorithm based on a classical-quantum hybrid neural network. In order to adapt the features of the Swin transformer and fully utilize the entanglement characteristics of QNN, we proposed an alternating layered ansatz (ALT)-based QNN, which successfully provides high trainability and rich expressibility. Extensive experiments on various generative model datasets with different training sample sizes indicate that with a small number of training samples, the hybrid neural network can achieve strong generalization, where the overall detecting accuracy is 2% higher than classical neural network. In addition, we verified the feasibility of a hybrid neural network on the “Wukong" 72-qubit superconducting quantum computer provided by OriginQ, showing similar accuracy to classical computer simulations, demonstrating that our model is feasible and effective in actual quantum computing systems.