DiffGait: Gait Representation Augmentation of mmWave Sensing Through a Latent Diffusion Model

User identification is a pivotal component of smart spaces, serving as a fundamental requirement of numerous smart applications, such as access control, precision management, personalized service, and so on. Compared with other solutions, millimeter-wave (mmWave) radar-based gait identification has the merits of being passive, noninvasive, and privacy-preserving, positioning it as the most promising approach for user identification. Yet, current neural network-based approaches still struggle with the challenge of collecting sufficient and diverse gait data during the training phase, which significantly relies on specific devices, spatiotemporal environments, and individual states. To this end, this article introduces DiffGait, a two-stage model for augmenting gait representations to enrich the training data samples for mmWave-based gait identification. In the initial stage, DiffGait utilizes a gait autoencoder and an adversarial discriminator to map extracted gait features into a low-dimensional latent space. Subsequently, a conditional latent diffusion model (LDM) is trained in the second stage, capable of generating latent representations of gaits from completely random noise inputs. In this way, our DiffGait can synthesize gait images by feeding the generated latent representations from the second stage into the decoder of the trained autoencoder from the first stage. Our extensive experimental results demonstrate the effectiveness of our approach in augmenting mmWave sensing data samples, particularly in improving both sample quantity and diversity, thereby offering a promising solution for data augmentation in wireless sensing.