Gang Zhou

Freezing of Gait (FoG) is one of the most debilitating symptoms of Parkinson's disease (PD), affecting over half of patients and severely impairing mobility and independence. Effective management requires both real-time FoG detection for reactive intervention and pre-FoG detection for proactive prevention before freezing fully manifests. However, existing methods rely on massive labeled data and require patient-specific calibration, restricting scalability and clinical deployment. We propose FM-FoG, a domain-specific foundation model designed for real-time FoG management without requiring individual patient-specific training. FM-FoG uses body-worn IMU sensors for gait monitoring, wearable vibrotactile actuators for delivering intervention cues, and a smartphone for real-time model inference. FM-FoG is pretrained in a self-supervised manner on approximately 73 hours of unlabeled IMU recordings from 170 subjects across diverse datasets, enabling generalizable motion representations that support two downstream tasks: real-time FoG detection and predictive pre-FoG detection. However, directly applying foundation models to wearable sensing presents two real-world challenges: (i) heterogeneous sensor configurations, and (ii) energy constraints in continuous operation. To address the first challenge, FM-FoG incorporates sensor-context embeddings that enable adaptive alignment across heterogeneous sensor locations. To address the second challenge, FM-FoG adopts an event-triggered activation strategy to limit inference to ambulatory periods and reduce unnecessary computation. Deployed on a Google Pixel 8a smartphone and evaluated under cross-patient protocols on our VCU FoG-IMU dataset of 23 PD patients with clinically induced FoG, FM-FoG achieves 98.5% F1 for FoG detection and > 85% F1 for 1-3s pre-FoG detection on completely unseen test subjects, and 96.9% and 93.8% respectively under leave-one-subject-out cross-validation, exceeding state-of-the-art models by up to 15%, while extending battery life by 43% at clinically representative activity levels with sub-20 ms latency, enabling practical, energy-efficient, and patient-independent FoG monitoring.

Recent advances in large language models (LLMs) create opportunities to expand access to mental healthcare. An emerging body of work focuses on LLM-based cognitive restructuring (CR), yet existing systems are often limited to generating alternative beliefs or engaging in short, scripted exchanges, offering little of the sustained interactivity required for effective restructuring. In this work, we present CRBot , an LLM-powered chatbot designed with mental health professionals to deliver CR. We conducted a two-week user study with 19 participants, complemented by expert review of conversation transcripts. Results show that CRBot can scaffold the core steps of CR, support self-reflection through Socratic questioning, and provide effective psychoeducation. At the same time, challenges arose, including rigidity, oversimplified questioning, pathologization of appropriate emotions, and misinterpretation of nuanced context. We conclude with design implications for building more interactive, flexible, and ethically robust LLM-enabled CR systems.

Graph embedding converts a graph data into a low dimensional space to preserve the original graph information. However, graph data can be reconstructed by malicious adversaries to train machine learning models from graph embeddings. This paper studies to what extent an adversary (without the original graph data) can recover the original graph data from graph embeddings. To quantify the original graph information leakage from graph embeddings, we develop a deep neural network model InferNet that can be used by adversaries to infer the original graph information from an adversary-accessible graph embedding database. More specifically, we propose the data-free reversed knowledge distillation technique to support the InferNet training even if the original graph dataset is absent. To ensure the performance of InferNet, we design two cycle-consistency loss functions to have an interactive training of InferNet over three series of datasets. To further enhance the performance of InferNet, we provide a joint training algorithm that simultaneously trains the pseudo-sample generator and InferNet, which significantly reduces the storage space. We evaluate the performance of InferNet on three datasets, and the intensive experiments demonstrate that InferNet can infer the original graph information from the graph embedding dataset with high accuracy.

Deep neural networks can be easily fooled by the adversarial example, which is a specially crafted example with subtle and intentional perturbations. A plethora of papers have proposed to use filters to effectively defend against adversarial example attacks. However, we demonstrate that the automatic filter-based defenses may not be reliable. In this article, we present URL2AED, Using a Reinforcement Learning scheme TO escape the automatic filter-based Adversarial Example Defenses. Specifically, URL2AED uses a specially crafted policy gradient reinforcement learning (RL) algorithm to generate adversarial examples (AEs) that can escape automatic filter-based AE defenses. In particular, we properly design reward functions in policy-gradient RL for targeted attacks and non-targeted attacks, respectively. Furthermore, we customize training algorithms to reduce the possible action space in policy-gradient RL to accelerate URL2AED training while still ensuring that URL2AED generates successful AEs. To demonstrate the performance of the proposed URL2AED, we conduct extensive experiments on three public datasets in terms of different perturbation degrees of parameter, different filter parameters, transferability, and time consumption. The experimental results show that URL2AED achieves high attack success rates for automatic filter-based defenses and good cross-model transferability.

Finger-vein biometrics has attracted significant attention in recent years, posing a challenge in extracting robust finger-vein patterns from raw images with limited prior knowledge. While deep learning-based models, particularly convolutional neural network (CNN)-based methods, exhibit substantial capacity for feature representation, they still suffer from certain drawbacks. First, CNN-based models focus on learning internal feature representations by minimizing the loss function of the reconstruction error. However, this strategy may lack optimal exploration of the parameter space, resulting in limited performance. Second, existing deep learning-based feature extraction approaches encode vein textures based on an output threshold of 0.5, potentially causing the omission of discriminant vein patterns. To address these drawbacks, we propose DRL-FEE, a Deep Reinforcement Learning-based Feature Extraction and Encoding for finger-vein verification in this paper. DRL-FEE comprises a DRL-based feature extraction approach for vein texture extraction and a DRL-based supervised encoding approach for vein feature encoding. Specifically, we design a CNN model incorporated into an RL framework to establish a DRL-based segmentation model vein feature extraction, utilizing the CNN to generate a policy for feature extraction. To further enhance performance, we develop a DRL-based supervised encoding approach with a transformer-based controller to search for an optimal threshold for vein feature encoding. The experimental results on three public finger-vein databases demonstrate that the proposed DRL-FEE outperforms state-of-the-art solutions, achieving the lowest verification errors with EERs of 3.13%, 2.86%, and 1.66%, respectively.

With the rapid development of the Internet of Things (IoTs) and mobile communications, mobile devices have become indispensable in our daily lives. Given the substantial amount of private information stored on these devices, the security of mobile devices has emerged as a significant concern for users. Different from conventional methods such as PINs, fingerprints, and face IDs, which authenticate users only during the initial login stage, continuous authentication ensures consistent verification while mobile devices are in use. Current continuous authentication methods require extensive data from a series of users for effective training. Nevertheless, it is challenging to collect sufficient amount of data within a limited time. In this paper, we propose CALL, an unsupervised sensor-based Continuous Authentication system with a Low-rank transformer using Learning-to-rank algorithms. The lightweight CALL is capable of providing both spatial and temporal features for end-to-end authentication. Specifically, CALL utilizes time series data from a legitimate user, collected by the accelerometer, gyroscope, and magnetometer sensors on smartphones, to train a pure one-dimensional autoencoder for spatial features and a shuffle low-rank Transformer (SLRT) for temporal features in the training phase. In the authentication phase, the trained pure one-dimensional autoencoder captures spatial features by reconstructing input data to obtain the reconstruction error, and SLRT captures temporal features by predicting a ranking vector that reveals the order of the shuffled feature sequence. The predicted ranking vector is then used to recover the shuffled sequence and the similarity between the frequency spectrum sequences of the recovered sequence and the original time series data is calculated. The reconstruction error and similarity are compared against pre-defined thresholds, and CALL authenticates a user as legitimate only if both values fall below their respective thresholds. Finally, we evaluate the performance of CALL on UCI_HAR, WISDM_HARB, and our dataset, and the extensive experiments illustrate that CALL reaches the best performance with 96.43%, 95.24% and 96.92% accuracy, and 4.28%, 4.76% and 3.86% EERs on the three datasets, outperforming state-of-the-art continuous authentication methods.