Zheng Zhao

Dr. Zheng Zhao

orcid

Research Focus & Metrics

Representative Publications

View All Publications
2 August 2024
Verified
Conference Paper
Open access
ZZZheng Zhao
Zheng Zhao
GGGuohua Gao
Guohua Gao
ZZZihan Zhang
Zihan Zhang
HHHouhua Huang
Houhua Huang
HLHongpeng Liu
Hongpeng Liu
The diversity of marine life poses a complex challenge for image recognition, particularly in the field of whale species, where there's a notable class imbalance. Some whale species have limited image samples, hindering the effectiveness of traditional deep learning methods. To tackle this, our study proposes a deep learning approach using the EfficientNet-V2 architecture, designed to enhance recognition precision and address the class imbalance. This method begins with preprocessing the input images to reduce noise and improve quality. It then employs a deep Convolutional Neural Network (CNN) to extract features from the images, which are aggregated using Generalized Mean Pooling (GeM) to form a rich feature representation. To optimize classification, the model incorporates ArcFace loss and Focal loss functions, which enhance feature distinctiveness and the recognition of difficult instances, respectively. The system includes classification and regression branches for species identification and feature localization. The model outputs prediction probabilities through a softmax function and evaluates performance using MAP@5. Notably, the EfficientNet-V2 model achieved a MAP@5 score of 0.856, surpassing the EfficientNet-B0 by 0.154.

21 August 2026 in Journal of King Saud University Computer and Information Sciences
Verified
Journal Article
Open access
PDF file
YCYuting Cao
Yuting Cao
ZZZheng Zhao
Zheng Zhao
JWJiekai Wu
Jiekai Wu
CLChia Hau Lee
Chia Hau Lee
QOQi Wei Oung
Qi Wei Oung
Hoi Leong LeeHoi Leong Lee
Hoi Leong Lee
Cooperative unmanned aerial vehicle (UAV) path planning is one of the core challenges in the field of autonomous systems, requiring the simultaneous guarantee of safety, efficiency, and multi-agent collaborative consistency in continuously dynamic obstacle environments. Existing methods either rely on hand-crafted heuristic rules and thus lack adaptability, or employ deep reinforcement learning (DRL) but suffer from low sample efficiency and sparse reward problems, making it difficult to achieve high task success rates in complex dynamic scenarios. This paper proposes SkyAgent—a lightweight large language model (LLM)-driven hierarchical reinforcement learning framework specifically designed for adaptive cooperative path planning of two UAVs. SkyAgent adopts Qwen2.5-1.5B (Q4_K_M quantization, approximately 1 GB VRAM) as the high-level semantic planner, performing fully offline inference at approximately 1,900 ms model forward-pass per call on a local GPU via the Ollama framework without any external API; employs Proximal Policy Optimization (PPO) as the low-level continuous action controller; and achieves efficient decoupled coordination of the two-tier policy through an asynchronous sub-goal injection mechanism (replanning every $$K=20$$ K = 20 steps). In a continuous two-dimensional simulation environment of $$100 \times 100$$ 100 × 100 m $$^2$$ 2 containing 8 static and 3 dynamic obstacles, SkyAgent is systematically evaluated against 7 baseline methods (Random, Greedy, APF, A*, PPO, SAC, OraclePID) over 30 random seeds. Experimental results demonstrate that SkyAgent achieves a mission success rate of 96.7%, comprehensively outperforming all baseline methods, improving by 16.7 percentage points over the vanilla PPO baseline, reducing collisions by 83%, shortening path length by 27 m, and improving sample efficiency approximately 3.3 $$\times $$ × (requiring only 55k training steps to reach the 80% success rate threshold). In the Dynamic-heavy scenario (30 obstacles), the success rate still reaches 90.9%, demonstrating significant generalization robustness. Six ablation experiments quantitatively validate the indispensability of the LLM planner, sub-goal guidance, sensor observations, cooperative reward, and reward shaping components. This work provides a feasible technical pathway and a reproducible evaluation benchmark for the collaborative deployment of lightweight LLMs and RL on GPU-equipped resource-constrained edge devices; we note that the present evaluation is confined to a two-dimensional, idealized simulation and that on-board deployment presupposes a GPU-class accelerator rather than an MCU-class micro-UAV.

8 September 2026 in Journal of King Saud University Computer and Information Sciences
Verified
Journal Article
Open access
PDF file
XYXianjun Yang
Xianjun Yang
ZZZheng Zhao
Zheng Zhao
LWLiPeng Wang
LiPeng Wang
ZCZhao Chen
Zhao Chen
Zero-day exploit (ZDE) attacks are among the most severe threats to critical infrastructures such as cloud computing, industrial control, smart grids, and connected vehicles, owing to their unknown, stealthy, and highly destructive nature. Traditional signature- and rule-based intrusion detection is largely ineffective against them, while existing deep learning methods struggle with global distribution modeling, local contextual dependencies, and low-sample generalization on high-dimensional, sparse, temporally correlated traffic and system-call data. This paper proposes VAE-AttNet, which fuses a Variational Autoencoder (VAE) with Multi-Head Self-Attention. Unlike prior hybrid designs that apply attention to raw features, VAE-AttNet confines self-attention to the VAE’s low-dimensional latent space and couples it with a joint anomaly score fusing reconstruction error and attention entropy. A VAE encoder maps behavioral features into a latent space via evidence-lower-bound maximization; self-attention then models context over the latent sequence, and a joint reconstruction–attention-entropy score with a dynamic 99th-percentile threshold identifies zero-day attacks. On CICIDS2017, UNSW-NB15, and ADFA-LD, VAE-AttNet reaches 97.83% accuracy, 96.42% F1, and 0.9871 AUC on CICIDS2017 — 2.69%–17.57% above eight baselines (Autoencoder, Isolation Forest, One-Class SVM, LSTM-AE, etc.) in relative terms, including a 2.59–3.30 percentage-point absolute gain over the strongest baseline LSTM-AE — with an average 8.56% F1 gain over LSTM-AE across six cross-dataset settings and F1 = 0.9023 using only 10% of training data. These results are scoped to the three evaluated datasets under a hold-out-attack-type protocol; validation on production-scale, adversarial traffic remains future work.

26 August 2026 in Research Square
Verified
Preprint
Open access
PDF file
ZZZheng Zhao
Zheng Zhao
In spatially resolved single-cell data, the analysis task is often the discovery of tissue regions—contiguous groups of cells that should share a representation precisely because they are co-located—yet graph contrastive learning, the dominant framework for learning representations on such graphs, constructs its training signal from two augmented views of the same node, pushing every node away from every other node. We examine this mismatch directly. Replacing instancediscrimination positives with spatial-adjacency positives inside a heterogeneous graph transformer, we train a three-layer encoder on a graph of 458,983 cells across nine human colonic sections, with five cell lineages and fifty typed relations built from Delaunay triangulation, defining each sampled edge between physically adjacent cells as a positive pair. The change raises Moran’s I from 0.633 to 0.756 (p < 0.001) and spatial coherence from 0.765 to 0.862 (p ≈ 0.03) relative to the strongest baseline, at half the runtime and half the memory of the two-view alternative, while no advantage is established on compactness or silhouette. Two methodological findings accompany this: a masked-reconstruction branch degrades all four metrics, and the CHAOS compactness metric varies by a factor of 2.4 across clustering initializations on a deterministic representation, so single-run values cannot support comparison between methods. Reproducible code is available at https://github.com/z1409498017-boop/SpaHetGCL.git.

28 September 2026 in Symmetry
Verified
Journal Article
Open access
PDF file
YQYifan QIN
Yifan QIN
ZZZheng Zhao
Zheng Zhao
Network threat detection in the Internet of Things must exploit typed relations between devices: a device with unremarkable flow statistics may still be compromised relative to its ownership, co-location, and social ties. Existing graph-based intrusion detectors discard relation type, while lightweight detectors compress tabular rather than graph models. This paper presents Light-SIoT-GAD, a lightweight relation-aware graph detector that treats typed relations as first-class input. The model learns one scalar weight per relation, shares a small basis across relation transforms so that each additional relation type costs only a handful of mixing coefficients rather than a full weight matrix, and combines supervised feature selection, bounded neighbour sampling, and 8-bit post-training quantisation for gateway deployment. A dual-track evaluation isolates typed aggregation on a real 16,216-device SIoT relation graph with injected anomalies and tests attack detection on three labelled NetFlow v2 corpora against classical, deep tabular, and graph baselines, including cross-corpus transfer and ablations. Light-SIoT-GAD matches or exceeds the strongest untyped graph baselines on all three corpora, reaching an AUPRC of 0.9986 with 73,962 parameters, substantially fewer than the unshared R-GCN; on the sparsest corpus a gradient-boosted tabular ensemble still ranks better, which bounds the claim to graphs that carry usable structure. Whether the gain comes from the relation semantics or merely from having per-relation parameters is tested directly: permuting the relation labels leaves the dense corpora unchanged to four decimals and costs 0.0083 AUPRC only on the sparsest one, so the typed advantage is real there and is a capacity effect elsewhere. On the social track, where the anomalies are injected under a stated protocol rather than observed, removing message passing collapses AUPRC from 0.9990 to 0.4869 under that same injection, isolating propagation over the relation graph as the source of the gain; this track measures whether relational structure carries signal, not accuracy against real SIoT attacks. The quantised model occupies 85.4 KB at 14.38 ms per 1000 edges on CPU. The learned relation gates are shown to be identified only up to a per-relation rescaling, and a leave-one-relation-out intervention finds no relation of the SIoT taxonomy to be load-bearing under the injection protocol, so the social track supports the typed parameterisation and not a ranking of the relations.