Q-Attentive DT-QFL: Adaptive Time-Symmetry in Quantum Federated Learning using Quantum Self-Attention

The unstoppable growth of the Industrial Internet of Things (IIoT) already has left the field of outright decentralization of data, provoking an urgent need in the creation of sensitive machine learning models that consider user privacy. The de facto standard towards this goal has been Federated Learning (FL), which allows edge devices to be trained without providing access to raw information. However, the dynamism of IIoT systems with Non-Independent and Identically Distributed (Non-IID) data is expected to worsen the performance of FL, resulting in such problems as model divergence and catastrophic forgetting. Quantum Federated Learning (QFL) solves these difficulties by transferring project data to high dimensional Hilbert spaces with Variational Quantum Circuits (VQCs), which is able to discover correlations that cannot be observed with classical networks. Among the latest inventions in this respect is the usage of TimeReversal Symmetry, another concept of physics which has been used to stabilize the learning process, based on matching the present model states with their preceding ones. The existing methods, such as Dual-Timeline QFL (DT-QFL) are flawed in the sense that they use hyperparameters to weight historical snapshots that are known beforehand. This rigidity may lead to slow convergence and too much communication overhead especially among networks. The present paper suggests a new model called Q-Attentive DT-QFL that implements a Quantum Self-Attention Mechanism (QSAM) in the time-reversal workflow. The model defines the relative values of past data at a certain point in time on its own based on dynamically computing attention scores between the current parameter trajectory and time reversed states. To support our proposal we present in depth theoretical analysis which illustrates convergence limits and reveals that our adaptive weighting methodology forms an important minimization of the variance of global updates. Scaling experiments on the Quantum MNIST dataset have shown that Q -Attentive DT -QFL has the highest accuracy in classification with 94.2 per cent accuracy and consumes forty percent less communication rounds compared with both fixed DT -QFL and classical baselines. b We also perform a detailed security review and certify the resistance of the framework to quantum noise (i.e. Depolarizing and Amplitude Damping ) and also compatibility with NearIntermediate-Scale Quantum (NISQ) hardware. In order to ensure transparency, we project the space of quantum features with SHAP values to explain the mechanisms of decision-making and use Principal Component Analysis(PCA).