Institution profile

University of Surrey

Academic institutioneurope · gb
Official website
Research library533linked papers
Opportunities0open roles
Selected work

Representative Papers

Efficient4D: Fast Dynamic 3D Object Generation from a Single-view Video

Jan 16, 2024

To address the challenges of missing 4D annotations and low end-to-end optimization efficiency in monocular video-based dynamic 3D reconstruction, this work proposes a two-stage decoupled paradigm: first generating multi-view temporally consistent images via a diffusion model, then driving 4D Gaussian Splatting for explicit reconstruction. We introduce an inconsistency-aware confidence-weighted loss and a lightweight Score Distillation Sampling (SDS) loss, significantly improving robustness under sparse-view conditions. Compared to Consistent4D, our method accelerates training tenfold (10 minutes vs. 120 minutes), enables real-time continuous trajectory rendering, and achieves state-of-the-art novel-view synthesis quality. To the best of our knowledge, this is the first work to organically integrate generative modeling with explicit 4D reconstruction, establishing a new paradigm for efficient, high-fidelity dynamic scene reconstruction.

42 citations5 influentialRead paper

Pruning Small Pre-Trained Weights Irreversibly and Monotonically Impairs"Difficult"Downstream Tasks in LLMs

Sep 29, 2023

This work challenges the prevailing assumption that small-magnitude weights in large language models (LLMs) are redundant, proposing instead the “Junk DNA Hypothesis”: low-magnitude weights encode essential knowledge for solving difficult downstream tasks. Method: We conduct systematic magnitude-based pruning—both structured and unstructured—alongside multi-granularity task difficulty quantification (e.g., reasoning depth, distribution shift, generalization gap), validated across model scales (7B–70B) and diverse benchmarks (MMLU, GSM8K, HumanEval). Contribution/Results: Pruning induces irreversible, monotonic performance degradation strictly correlated with task difficulty—degradation persists even after extensive fine-tuning—whereas quantization exhibits no such effect. This is the first study to empirically establish the functional necessity of small-magnitude weights from a task-difficulty perspective. We further propose novel, quantifiable cross-task difficulty metrics and demonstrate a strong negative correlation between optimal pruning ratio and task difficulty.

7 citations1 influentialRead paper

Unified Source-Free Domain Adaptation

Mar 12, 2024arXiv.org

Existing source-free domain adaptation (SFDA) methods are constrained to specific settings—e.g., closed-set, open-set, biased-set, or generalized SFDA—and rely on target-domain priors, limiting their applicability and theoretical grounding. Method: This work introduces Unified SFDA, the first formal problem formulation of SFDA that requires neither source data nor target-domain prior knowledge. From a causal perspective, it models the generative relationship between latent variables and decisions, proposing the Latent Causal Factor Discovery (LCFD) framework. LCFD integrates vision-language pretrained models (e.g., CLIP) with a causally motivated information bottleneck objective to achieve theoretically guaranteed representation disentanglement. Contribution/Results: Unified SFDA establishes a general, prior-free SFDA paradigm. It achieves state-of-the-art performance across all major SFDA benchmarks and significantly improves out-of-distribution generalization, demonstrating robustness to unseen domain shifts without access to source data or target annotations.

6 citations1 influentialRead paper

Diffusion-based Generative Multicasting with Intent-aware Semantic Decomposition

Nov 04, 2024arXiv.org

To address low-latency and heterogeneous semantic requirements in multi-user semantic communications for future wireless networks, this paper proposes an intent-aware generative semantic multicast framework. The transmitter decomposes the source signal according to each user’s semantic intent, transmitting only the intended semantic classes while broadcasting a lightweight shared semantic graph; users then collaboratively reconstruct non-intended classes locally using pre-trained diffusion models. This work pioneers the integration of generative diffusion models (GDMs) into semantic multicast, enabling intent-driven semantic decomposition and generative reconstruction. We further design a communication-computation co-optimized, per-class adaptive parameter allocation mechanism that jointly optimizes transmit power, coding rate, and model inference overhead. Experimental results demonstrate that, compared to conventional non-generative and intent-agnostic baselines, the proposed framework significantly reduces end-to-end latency, improves spectral efficiency, and enhances privacy protection for non-intended semantic content.

2 citationsRead paper

KPI Poisoning: An Attack in Open RAN Near Real-Time Control Loop

May 08, 2025

This work addresses KPI poisoning attacks in the Open RAN near-real-time (Near-RT) control loop, induced by E2 interface traffic spoofing or node hijacking—attacks systematically formalized here for the first time. To ensure closed-loop control security, we propose a lightweight temporal anomaly detection paradigm. Our method employs an LSTM-based model for KPI stream anomaly detection, trained on realistic E2-interface KPI report simulations augmented with a controllable poisoning injection mechanism. The design guarantees sub-millisecond detection latency while enhancing robustness against adversarial perturbations. Experimental evaluation demonstrates a substantial improvement in detection rate—from 62% to 99%—and identifies two key determinants of detection performance: the degree of poisoning amplification and the length of the input time series. This work delivers a deployable, theoretically grounded detection framework for securing the Open RAN control plane.

1 citationsRead paper
Recent publications

Latest Papers