Research

Industrial Foundation Models (IFMs) represent a transformative shift in smart manufacturing, offering unprecedented multi-task generalization through multimodal data integration. However, the transition from general-purpose AI to factory-floor deployment is hindered by stringent requirements for precision and the lack of standardized benchmarks. While existing surveys focus on high-level applications, this paper provides a granular technical roadmap centered on the data-centric and metric-centric challenges of the field. We deliver an exhaustive analysis of publicly available datasets and expose the critical misalignment between standard machine learning scores and practical manufacturing constraints. By consolidating these technical pillars, this work equips researchers with the essential resources to develop trustworthy, production-ready IFMs for the next generation of industrial intelligence.

Third-generation Fully Homomorphic Encryption (FHE), particularly the FHEW/TFHE schemes, is recognized for its balanced security requirements, small parameters, and low memory usage, though the current methods in the scenarios of Deep Neural Network (DNN) inference still have high computational costs, limiting the practical applicability. This work demonstrates how to improve practicality  of  the third-generation technologies  for   DNN tasks while preserving its key advantages. Our work focuses on two main contributions. First, we developed a computational architecture called FHE-Neuron, which reconfigures the parameters and bootstrapping structure of traditional FHEW/TFHE Boolean operations. This architecture significantly reducing the cost of encrypted DNN inference by dynamically switching the precision of encrypted data during computation---using high precision for cost-effective linear operations and low precision for computationally expensive nonlinear operations. Second, we introduced an FHE-aware Quantization and Fine-tuning framework that optimizes model parameters to align with FHE-Neuron’s constraints, ensuring high accuracy in encrypted inference.  We validate our approach on various neural network models across several computing platforms. In our experiments, our method achieves one-image inference time on average 4.5 milliseconds for MNIST and 17 milliseconds for Fashion MNIST, achieving accuracy rates of 96.52% and 88.57% respectively. For the CIFAR-10 dataset, our system completes one image inference in 30 seconds with a 90.5% accuracy rate.

Most Federated Learning (FL) systems are built upon a strong assumption of trust---clients fully trust the centralized moderator, which might not be feasible in practice. This work aims to mitigate the assumption by using appropriate cryptographic tools. Particularly, we examine various scenarios with different trust demands in FL, and then design the corresponding practical protocols with lightweight cryptographic tools. Three solutions for secure and trustworthy aggregation are proposed with increasing sophistication: (1) a single verifiable moderator, (2) a single secure and verifiable moderator, and (3) multiple secure and verifiable moderators, which can handle adversarial behaviors with different levels. We evaluate the performances of all our proposed protocols on the test accuracy and the training time, showing that our protocols maintain the accuracy with time overhead from 30% to 156% depending on the secure and trustworthy levels. The protocols can be deployed in many practical FL settings with appropriate optimizations.

Flow matching and diffusion models have shown impressive results in text-to-image generation, producing photorealistic images through an iterative denoising process. A common strategy to speed up synthesis is to perform early denoising at lower resolutions. However, traditional methods that downscale and upscale in pixel space often introduce artifacts and distortions. These issues arise when the upscaled images are re-encoded into the latent space, leading to degraded final image quality. To address this, we propose Latent Space Scaling Generation(LSSGen), a framework that performs resolution scaling directly in the latent space using a lightweight latent upsampler. Without altering the Transformer or U-Net architecture, LSSGen improves both efficiency and visual quality while supporting flexible multi-resolution generation. Our comprehensive evaluation covering text-image alignment and perceptual quality shows that LSSGen significantly outperforms conventional scaling approaches. When generating 10242 images at similar speeds, it achieves up to 246% TOPIQ score improvement 

Continual learning (CL) seeks to mitigate catastrophic forgetting when models are trained with sequential tasks. A common approach, experience replay (ER), stores past exemplars but only sparsely approximates the data distribution, yielding fragile and oversimplified decision boundaries. We address this limitation by introducing Support Boundary Data (SBD), generated via differential-privacy-inspired noise into latent features to create boundary-adjacent representations that implicitly regularize decision boundaries. Building on this idea, we propose Experience Blending (EB), a framework that jointly trains on exemplars and SBD through a dual-model aggregation strategy. EB has two components: (1) latent-space noise injection to generate support boundary data, and (2) end-to-end training that jointly leverages exemplars and SBD. Unlike standard experience replay, SBD enriches the feature space near decision boundaries, leading to more stable and robust continual learning. Extensive experiments on CIFAR-10, CIFAR-100, Tiny ImageNet, and ImageNet-1k demonstrate consistent accuracy improvements of 10%, 6%, 14%, and 9%, respectively.

Immersive video streaming technologies improve Virtual Reality (VR) user experience by providing users more intuitive ways to move in simulated worlds, e.g., with 6 Degree-of-Freedom (6DoF) interaction mode. A naive method to achieve 6DoF is deploying cameras at numerous different positions and orientations that may be required based on users' movement, which unfortunately is expensive, tedious, and inefficient. A better solution for realizing 6DoF interactions is to synthesize target views on-the-fly from a limited number of source views. While such view synthesis is enabled by the recent Test Model for Immersive Video (TMIV) codec, TMIV dictates manually-composed configurations, which cannot exercise the tradeoff among video quality, decoding time, and bandwidth consumption. In this article, we study the limitation of TMIV and solve its configuration optimization problem by searching for the optimal configuration in a huge configuration space. We first identify the critical parameters in the TMIV configurations. Then, we introduce two Neural Network (NN)-based algorithms from two heterogeneous aspects: (i) a Convolutional Neural Network (CNN) algorithm solving a regression problem and (ii) a Deep Reinforcement Learning (DRL) algorithm solving a decision making problem, respectively. We conduct both objective and subjective experiments to evaluate the CNN and DRL algorithms on two diverse datasets: an equirectangular and a perspective projection datasets. The objective evaluations reveal that both algorithms significantly outperform the default configurations. In particular, with the equirectangular (perspective) projection dataset, the proposed algorithms only require 95% (23%) decoding time, stream 79% (23%) views, and improve the utility by 6% (73%)on average. The subjective evaluations con firm the proposed algorithms consume less resources while achieving comparable Quality of Experience (QoE) than the default and the optimal TMIV configurations.

Although live video communication is widely used, it is generally less engaging than face-to-face communication because of limitations on social, emotional, and haptic feedback. Missing eye contact is one such problem caused by the physical deviation between the screen and camera on a device. Manipulating video frames to correct eye gaze is a solution to this problem. In this paper, we introduce a system to rotate the eyeball of a local participant before the video frame is sent to the remote side. It adopts a warping-based convolutional neural network to relocate pixels in eye regions. To improve visual quality, we minimize the L2 distance between the ground truths and warped eyes. We also present several newly designed loss functions to help network training. These new loss functions are designed to preserve the shape of eye structures and minimize color changes around the periphery of eye regions. To evaluate the presented network and loss functions, we objectively and subjectively compared results generated by our system and the state-of-the-art, DeepWarp, in relation to two datasets. The experimental results demonstrated the effectiveness of our system. In addition, we showed that our system can perform eye gaze correction in real time on a consumer-level laptop. Because of the quality and efficiency of the system, gaze correction by postprocessing through this system is a feasible solution to the problem of missing eye contact in video communication.

Data from multi-modal sensors, such as RGB cameras, thermal cameras, microphones, and mmWave radars, have gradually been adopted in various classification problems for better accuracy. Some sensors, like RGB cameras and microphones, however, capture privacy-invasive data, which are less likely to be used in centralized learning. Although the Federated Learning (FL) paradigm frees clients from sharing their sensor data, doing so results in reduced classification accuracy and increased training time. In this article, we introduce a novel Heterogeneous Privacy Federated Learning (HPFL) paradigm to better capitalize on the less privacy-invasive sensor data, such as thermal images and mmWave point clouds, by uploading them to the server for closing the performance gap between FL and centralized learning. HPFL not only allows clients to keep the more privacy-invasive sensor data private, such as RGB images and human voices, but also gives each client total freedom to define the levels of their privacy concern on individual sensor modalities. For example, more sensitive users may prefer to keep their thermal images private, while others do not mind sharing these images. We carry out extensive experiments to evaluate the HPFL paradigm using two representative classification problems: semantic segmentation and emotion recognition. Several key findings demonstrate the merits of HPFL: (i) compared to FedAvg, it improves foreground accuracy by 18.20% in semantic segmentation and boosts the F1-score by 4.20% in emotion recognition, (ii) with heterogeneous privacy concern levels, it achieves an even larger F1-score improvement of 6.17–16.05% in emotion recognition, and (iii) it also outperforms the state-of-the-art FL approaches by 12.04–17.70% in foreground accuracy and 2.54–4.10% in F1-score.

Foveated rendering leverages human visual system to increase video quality under limited computing resources for Virtual Reality (VR). More specifically, it increases the frame rate and the video quality of the foveal vision via lowering the resolution of the peripheral vision. Optimizing foveated rendering systems is, however, not an easy task, because there are numerous parameters that need to be carefully chosen, such as the number of layers, the eccentricity degrees, and the resolution of the peripheral region. Furthermore, there is no standard and efficient way to evaluate the Quality of Experiment (QoE) of foveated rendering systems. In this paper, we propose a framework to compare the performance of different subjective assessment methods on foveated rendering systems. We consider two performance metrics: efficiency and consistency, using the perceptual ratio, which is the probability of the foveated rendering is perceivable by users. A regression model is proposed to model the relationship between the human perceived quality and foveated rendering parameters. Our comprehensive study and analysis reveal several insights: 1) there is no absolute superior subjective assessment method, 2) subjects need to make more observations to confirm the foveated rendering is imperceptible than perceptible, 3) subjects barely notice the foveated rendering with an eccentricity degree of 7.5◦+ and peripheral region of a resolution of 540p+, and 4) QoE levels are highly dependent on the individuals and scenes. Our findings are crucial for optimizing the foveated rendering systems for future VR applications.