With the rapid advancement of AI-generated videos, distinguishing synthetic content from genuine human-driven content has become increasingly difficult, threatening the integrity of human authenticity and creative expression. In this context, Cybernetic Avatar (CA) is introduced as a digital entity that mirrors a remote operator’s facial expressions, gestures, and speech in virtual environments, posing new challenges for secure identity verification. A critical threat emerges when unauthorized users manipulate a CA, potentially deceiving both systems and human observers. This paper addresses the CA Authentication problem, which seeks to verify the true teleoperator behind a CA video despite the CA’s mutable appearance and expressive behaviors. More specifically, we propose a robust CA authentication framework that leverages spatio-temporal facial behavior captured from the CA video to authenticate the legitimate teleoperator. To effectively learn identity-sensitive motion patterns (signature), we develop a Behavior Signature Decoder Graph Convolutional Network (BSDec-GCN) that constructs a constrained spatio-temporal graph to amplify identity-specific dynamics and suppress inter-user ambiguity. Furthermore, we introduce a dual landmark and graph-level losses that boost discrimination. The comprehensive experiments and the thorough ablation studies demonstrate the reliability of the proposed framework with a competitive performance against the existing baseline methods. To the best of our knowledge, this work presents the first graph-based learning approach tailored for Cybernetic Avatar authentication, opening a new direction for securing virtual identity in the era of AI-mediated communication.

Current face de-identification methods often struggle to balance robust privacy protection with preserving image utility. While existing approaches effectively obscure identity, they frequently degrade visual quality, limiting their practical applicability. To address this challenge, we propose a robust reconstruction and de-identification framework (RaD) that leverages the 3D Morphable Model (3DMM) explicit representation. We first utilize a CNN encoder to predict the disentangled 3DMM coefficients, enabling a coarse reconstruction via a differentiable renderer. To enrich facial detail beyond the 3DMM’s topology and statistical priors, we introduce a personalized albedo generator (PAG) that adds fine texture details. For de-identification, we then apply a transformer-based identity protector (IP) to manipulate the 3DMM shape and texture parameters in order to conceal identity-sensitive features while preserving the photorealism of the reconstructed images. Finally, an Image Enhancement Module (IEM) refines the outputs, removing any artifacts and further enhancing visual quality. Extensive experiments on multiple benchmarks demonstrate that our framework outperforms state-of-the-art methods both quantitatively and qualitatively, making it well-suited for applications that require realistic reconstructions along with privacy preservation without compromising image quality.

This paper presents an online human model-based framework for gait-based age and gender estimation from a sequence of monocular frames. More specifically, we fine-tune a human mesh recovery model (i.e., HMR) to estimate the shape and pose parameters of a predefined 3D human model (i.e., SMPL). We then utilize the estimated parameters to predict the age and gender of the walking subject. To make the age and gender estimation task more favorable for real-time applications, we consider estimating the corresponding probability distributions of age and gender, which preserve the prediction uncertainty. Experiments on the world’s largest multi-view gait age and gender estimation dataset showed the superiority of the proposed method compared to the existing appearance-based baseline. We implement online standalone and client-server systems based on the proposed framework to demonstrate the performance of real-time estimation. We further propose a geometric correction step to the input gait sequence for a more generalization capability of the online system.

In this paper, we describe an uncertainty-aware estimation framework for gait relative attributes. We specifically design a two-stream network model that takes a pair of gait videos as input. It then outputs a corresponding pair of Gaussian distributions of gait absolute attribute scores and annotator-dependent gait relative attribute label distributions. Moreover, we propose a differentiable annotator-independent uncertainty layer to estimate the gait relative attribute score distribution from the absolute distributions then map it to a relative attribute label distribution using the computation of cumulative distribution functions. Furthermore, we propose another annotator-dependent uncertainty layer to estimate the uncertainty on the gait relative attribute labels in terms of a set of trainable transition matrices. Finally, we design a joint loss function on the relative attribute label distribution to learn the model parameters. Experiments on two gait relative attribute datasets demonstrated the effectiveness of the proposed method against baselines in quantitative and qualitative evaluations.

This paper describes a method for estimating gait relative attribute distributions. Existing datasets for gait relative attributes have only three-grade annotations, which cannot be represented in the form of distributions. Thus, we first create a dataset with seven-grade annotations for five gait relative attributes (i.e., beautiful, graceful, cheerful, imposing, and relaxed). Second, we design a deep neural network to handle gait relative attribute distributions. Although the ground-truth (i.e., annotation) is given in a relative (or pairwise) manner with some degree of uncertainty (i.e., inconsistency among multiple annotators), it is desirable for the system to output an absolute attribute distribution for each gait input. Therefore, we develop a model that converts a pair of absolute attribute distributions into a relative attribute distribution. More specifically, we formulate the conversion as a transportation process from one absolute attribute distribution to the other, then derive a differentiable model that determines the trade-off between optimal transport and uniform transport. Finally, we learn the network parameters by minimizing the dissimilarity between the estimated and ground-truth distributions through the Kullback–Leibler divergence and the expectation dissimilarity. Experimental results show that the proposed method successfully estimates both absolute and relative attribute distributions.

This paper presents a deep learning-based method to estimate gait attributes (e.g., stately, cool, relax, etc.). Similarly to the existing studies on relative attribute, human perception-based annotations on the gait attributes are given to pairs of gait videos (i.e., the first one is better, tie, and the second one is better), and the relative annotations are utilized to train a ranking model of the gait attribute. More specifically, we design a Siamese (i.e., two-stream) network which takes a pair of gait inputs and output gait attribute score for each. We then introduce a suitable loss function called a signed contrastive loss to train the network parameters with the relative annotation. Unlike the existing loss functions for learning to rank does not inherit a nice property of a quadratic contrastive loss, the proposed signed quadratic contrastive loss function inherits the nice property. The quantitative evaluation results reveal that the proposed method shows better or comparable accuracies of relative attribute prediction against the baseline methods.