QPQianqian Pan

Qianqian Pan

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Qianqian Pan has a total of 39 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Secure Federated Data Augmentation Empowered Efficient THz Semantic Communication for Agentic AI Collaborative Networking

    by Zihong Li, Jinhao Zhou, Qianqian Pan, Jun Wu

    Agentic artificial intelligence (AI) collaborative networking and secure federated learning are emerging as critical enablers for unmanned swarms with the advent of 6G and the proliferation of AI-enabled IoT devices, where autonomous driving and cooperative tasks require fast, reliable, and energy-efficient communication and learning. However, massive multimodal sensor data, the limited bandwidth of conventional mmWave V2X links, heterogeneous local datasets, privacy concerns over sensitive observations, and the strict energy budgets of UAVs, UUVs, and small robots create major bottlenecks for traditional federated learning. To address these challenges, we propose a cross-layer network framework integrating 6G THz-enabled efficient semantic communication with secure federated data augmentation for agentic AI collaboration. Semantic compression over short-range THz links reduces per-round transmitted payloads by extracting only task-relevant features, while semantic prototype-based data augmentation accelerates FL convergence by enriching non-IID local datasets without additional communication overhead; secure aggregation (SecAgg) further ensures that individual model updates are never exposed to the aggregation server. This co-design simultaneously reduces per-round transmission cost and the total number of rounds required for convergence, yielding a compound reduction in cumulative communication energy. Experimental results demonstrate that the proposed system significantly outperforms a conventional mmWave-based federated learning baseline in transmission latency, tail-latency reliability, learning convergence speed, global model accuracy, and cumulative communication energy, validating its effectiveness for energy-efficient and privacy-preserving agentic AI collaboration.


  • Device-Bind Key-Storageless IP Protection for Cloud-AI Models via Permutational Diffusional PUF

    by Qianqian Pan, Mianxiong Dong, Kaoru Ota, Jun Wu

    With powerful cloud computing capacities, machine learning as a service (MLaaS) framework provides intelligent services and well-trained artificial intelligence (AI) models by clouds for resource-constrained end devices. However, there are AI model leakage and illegal abuse issues during model transmission and deployment. Existing mainstream protection methods face the following problems: (i) The watermark-based methods only provide passive verification afterward rather than active protection. (ii) Encryption-based methods are low efficiency in computation and low security in key storage. (iii) Existing methods are not device-bind without the ability to avoid illegal abuse issues. To deal with these problems, we propose a device-bind and key-storageless cloud-AI model intellectual property (IP) protection mechanism. First, a physical unclonable function (PUF) and permute-diffusion encryption empowered cloud-AI model protection framework is proposed, including the PUF-based secret key generation and the geometric-value transformation based weights encryption. Second, we design an Anderson PUF based key generation protocol to generate device-bind robust secret keys. Third, a permutation and diffusion-based intelligent model weights encryption/decryption method is proposed for effective cloud-AI model protection, where chaos theory is utilized to convert PUF-based secret keys to encryption/decryption keys. Finally, experimental evaluation demonstrates the effectiveness and reliability of the proposed intelligent model IP protection mechanism.


  • Toward Effective and Sustainable Traffic Management With Network Scaling Configuration: A Federated FSL Approach

    by Qianqian Pan, Akihiro Nakao

    In 5G and beyond networks, the highly dynamic and time-varying environment leads to multiple types of network function (NF) errors. NF scaling technology is capable of addressing this problem by scaling up/down NFs adaptively according to real-time network traffic. Existing NF scaling approaches face the following issues: 1) Threshold-based approaches adjust NFs when traffic loads exceed/fall preset thresholds, which is passive and has a long response time. 2) Forecasting-based approaches suffer from inaccurate predictions, insufficient training data, and difficulty adapting to all dynamic and varying network environments. 3) Reinforcement learning-based approaches improve robustness, but still struggle with the problem of insufficient data. To address the above issues and challenges, we propose a federated traffic prediction-based NF scaling mechanism. First, the federated traffic flow prediction and function configuration framework is designed, including fine-grained federated traffic prediction and intelligent NF configuration. Second, we propose a federated few-shot learning-driven traffic flow prediction scheme, where Siamese FSL networks are constructed to forecast the traffic flow of new areas with few samples. Third, we devise a deep reinforcement learning-driven online NF configuration scheme to select real-time NF scaling strategies for efficient and sustainable networks. Finally, experimental results demonstrate that our proposed federated traffic flow prediction-based NF scaling mechanism achieves better performance than baselines.


  • Program Interoperable Large Language Model Software Testing Scheme: A Case Study on JavaScript Engine Fuzzing

    by Daipeng Cao, Yang Hong, Qianqian Pan, Jun Wu

    Large language models (LLMs) have cultivated impressive semantics capabilities and expert knowledge from their vast pre-training corpora, especially showing prospects in automated software testing. However, LLMs are designed for human interaction, which poses the following challenges when interacting with programs for testing: 1) LLMs cannot communicate directly with programs, and there is no existing paradigm to establish interaction between them. 2) Existing evaluation methods are unable to assess the quality of LLMgenerated tests during software testing. 3) Current LLM-guided testing generation cannot be optimized in real time, resulting in low testing efficiency. To address these challenges, we present PILLM, a program interoperable LLM scheme. First, we designed a prompt mechanism for interactive program testing based on the source code semantics and expert knowledge from the LLM. Second, we proposed an evaluation mechanism for the PILLM's test code generation, thus obtaining test seeds that are semantically related to the corresponding source code. Third, PILLM optimizes the next generated tests based on the coverage and source code execution information obtained in the program execution. In the 24-hour running experiment, PILLM improved the coverage by 45.7% and 14.9% compared to Fuzz4all and Fuzzilli respectively. PILLM proves its effectiveness by finding four new real-world bugs in the JavaScript engine. We have released the source code of PILLM as open source on Github.


  • Incentive Distributed Knowledge Graph Market for Generative Artificial Intelligence in IoT

    by Guozhi Hao, Qianqian Pan, Jun Wu

    Generative artificial intelligence (GAI) models are pre-trained using extensive public data. However, in the Internet of Things (IoT) domain, distributed and heterogeneous data from end devices lacks contextual relations, which impairs IoT domain GAI training efficiency and leads to inaccurate applications. Nowadays, knowledge graph (KG) offers an effective solution for enhanced performance and interpretability of GAI. Nevertheless, specific data collection and KG creation in IoT scenarios still pose quality and cost challenges for GAI developers. Therefore, designing an effective IoT KG collection and trading strategy is needed to provide reliable knowledge support for IoT GAI from professional data providers. Currently, establishing fair KG pricing, the effective construction of knowledge relations in edge IoT scenarios and preventing data providers from accessing private information remain significant open issues for IoT KG trading. To address these issues, we propose an incentive distributed IoT knowledge market framework to facilitate effective and secure KG trading for IoT GAI. Specifically, first, we design a utility incentive and GAI demand-driven KG pricing strategy, which establishes a three-layer game with trading participants and KG embedding utility function to obtain fair pricing. Second, we devise a smart contract based distributed knowledge aggregation method, which provides collaborative IoT KG relation creation. Third, we propose a privacy-preserving KG construction scheme via homomorphic encryption that achieves consensus of encrypted KG to prevent knowledge providers from accessing IoT privacy. Finally, experimental results of real KG verify the KG-enhanced GAI and proposed market availability.


  • Large-Scale Mean-Field Federated Learning for Detection and Defense: A Byzantine Robustness Approach in IoT

    by Kangkang Sun, Lizheng Liu, Qianqian Pan, Jianhua Li, Jun Wu

    We present a three-part sensing system to measure the acquisition of foraging proficiency in wild tree squirrels. The first component is the eNut: a 3Dprinted enclosure in the size and shape of a large food item, such as a walnut. The eNut contains an inertial measurement unit (IMU), consisting of an accelerometer and a gyroscope, along with a capsule containing a food reward (e.g. chopped nuts), motivating the squirrel to manipulate the eNut to open the reward capsule. The second component is a radio-frequency identification (RFID) gate that can detect the presence of a known individual. The reader generates time-stamped labels for the analysis of simultaneous video recordings of the squirrel movements. Finally, the third component is a data collection pipeline, which aggregates the sensor data in a cloud backend server. The design process culminated in a fully-functional prototype system tested under three field conditions: two outdoor sites used for juvenile squirrel rehabilitation and a third site with free-ranging adult squirrels. The accuracy of the collected data is assessed by leveraging lab settings in which acceleration and rotation were checked against reference values produced by a testing infrastructure, custom-built for this application. Finally, we present an inventory of food item manipulation movements (rotation, shaking) that are identified and quantified with the eNut system. This system allows a level of granular analyses of foraging decisions that was not feasible with prior technology, which would have important translational value in paradigms of animal behavior and wildlife rehabilitation.


  • O-RAN-Based Digital Twin Function Virtualization for Sustainable IoV Service Response: An Asynchronous Hierarchical Reinforcement Learning Approach

    by Yihang Tao, Jun Wu, Qianqian Pan, Ali Kashif Bashir, Marwan Omar

    Digital Twin for Vehicular Networks (DTVN) continuously simulates and optimizes vehicle behaviors to support emerging 6G Internet-of-Vehicle (IoV) applications such as DT-assisted autonomous driving. To meet Quality of Service (QoS), resource scheduling for distributed vehicle DTs is carried out. However, existing works mainly respond to service demand based on one-to-one DT synchronization and computation offloading, which limits the service response quality and is not sustainable. Meanwhile, twin objects need to be frequently transferred at edges in parallel with the moving vehicles, the IoV service demand response under high-dynamic DT resource distribution is challenging. In this paper, a novel digital twin function virtualization (DTFV) architecture based on Open Radio Access Networks (O-RAN) is proposed. In DTFV, multiple vehicle DTs following one-to-one synchronization are decoupled and reorganized as a Virtualized Digital Twin (VDT) following dissemination-based synchronization for dynamic service response, without needs for offloading service to additional edge devices. Besides, to optimize the overall IoV service response profit, we propose an asynchronous hierarchical reinforcement learning (AHRL)-based DTFV resource scheduling scheme to find optimal VDT orchestration and synchronization strategies. Finally, experimental results show our scheme achieves 8.48% higher service response profit and 6.8% lower VDT synchronization delay over the best baseline scheme.