A Mouse Dynamics Authentication System With a Recurrence Plot Image Representation and a Vision Transformer Framework

In this paper, we propose a system that verifies the authenticity of users based on the manner in which they operate a computer mouse. To begin with, we introduce a recurrence plot representation for encoding the information available in the mouse dynamics. Two image representation variants are suggested, namely the symmetric and asymmetric recurrence plots. Another noteworthy contribution is a modified vision transformer architecture for this task that incorporates key adjustments such as the removal of class token and positional embeddings. Rather, we facilitate a local pattern classification by considering the use of feature aggregation strategy for decision making. Additionally, we incorporate an efficient attention mechanism within the transformer encoder, that reduces both computational and memory complexity by simplifying the attention process. To further boost model performance, we integrate the Gradient Harmonizing Mechanism with binary cross-entropy loss, which dynamically adjusts the loss function based on gradient magnitudes. The proposed system is evaluated on three publicly available datasets, and the results obtained are at par to state-of-the-art methods. To the best of our knowledge, the present proposal is the first of its kind to introduce the utility of recurrence plots in a modified transformer framework.