Pith. sign in

REVIEW 4 major objections 5 minor 54 references

Secure Resource Allocation via Constrained Deep Reinforcement Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read An action-constrained deep Q-network is claimed to cut serverless edge-cloud system costs by up to 40% and to improve energy efficiency by 41.5%.

desk verdict The paper's central novelty claim doesn't survive its own equations: the 'action constraint' is a fixed penalty on aggregate delay, and the experiments don't show that the resulting schedules are even feasible. read the letter →

arxiv 2501.11557 v1 pith:RYPI6PGJ submitted 2025-01-20 cs.LG

Jianfei Sun , Qiang Gao , Cong Wu , Yuxian Li , Jiacheng Wang , Dusit Niyato This is my paper · ORCID
classification cs.LG
keywords action-constraineddeepreinforcementlearningserverlessmulti-cloudedgecomputingresourceallocationtaskoffloadingsecurityoverheadQ-networkenergyefficiencydelayconstraints
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Serverless multi-cloud edge computing must decide where to run each IoT task while meeting per-task deadlines, respecting node capacities, and paying for security overhead. The paper argues that standard deep reinforcement learning handles these constraints poorly because it folds penalties into the reward, producing suboptimal policies. It proposes SARMTO, built around an action-constrained deep Q-network (AC-DQN), which injects a penalty into the Q-value of any action whose total delay exceeds a limit, thereby shaping the learned value function to avoid infeasible actions. In simulations with up to 1000 tasks and data sizes up to 5 GB, SARMTO consistently beats five baselines, reporting up to 40% lower system cost and 41.5% better energy efficiency. A sympathetic reading takes these as strong simulated gains; the load-bearing question is whether the action-constraint mechanism truly enforces the deadlines and capacities it targets.

What carries the argument

The carrying mechanism is the action constraint function $f_{\mathrm{constraint}}(a_j) = -\lambda$ if $T_{\mathrm{total}}(a_j) > T_{\max}$ else $0$ (Eq. 7), with $\lambda$ a large positive constant such as 1000. It is added to the Q-value during action selection (Algorithm 2, line 7) and inside the target Q-value calculation (line 17), so that infeasible actions receive a permanently depressed Q-value and the constraint is propagated through temporal-difference learning. This turns a hard feasibility check into a discontinuous correction of the action-value landscape, which the paper argues avoids the suboptimal equilibria of reward-penalty approaches. The Q-network uses a dueling architecture with prioritized experience replay and a target network; the security overhead equations (Eqs. 5, 6, 9, 10) quantify how encryption, decryption, and hashing cycles enter the delay and energy terms of the reward.

What would settle it

Run the trained AC-DQN policy in the simulator and count the fraction of actions whose total delay exceeds $T_{\max}$ or whose assigned tasks exceed a node's capacity $f_k$; if this fraction is not zero (or not dramatically lower than the unconstrained DQN baseline), the action-constraint mechanism is not enforcing the constraints it targets.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that a constraint term added directly to the Q-value function, rather than to the reward, lets deep reinforcement learning respect hard scheduling constraints in serverless multi-cloud edge environments while still optimizing a weighted delay-energy cost. The AC-DQN algorithm defines $f_{\mathrm{constraint}}(a) = -\lambda$ when the total delay of action $a$ exceeds $T_{\max}$ and zero otherwise (Eq. 7), adds this term when selecting actions and in the target $y = r + \gamma \max_{a'}\big(Q(s',a';\theta^-) + f_{\mathrm{constraint}}(a')\big)$, and trains the Q-network with the usual temporal-difference loss. The authors state this embeds constraints into the learned value function itself, contrasting with methods that only filter actions at selection time. Extensive simulations report that SARMTO achieves the lowest system cost across task counts, data sizes, and MEC capacities, with specific margins such as 23.6% over DQN at 1000 tasks, 26.3% at 5 GB data size, and up to 34.7% at low MEC capacity, plus 41.5% lower energy than CO at 5 GB. The framework also includes an adaptive security mechanism (RSA encryption and MD5 hashing with modeled overhead) and an adaptive $\epsilon$-greedy exploration schedule, but the distinguishing innovation is the action-constraint term.

Load-bearing premise

The single fixed penalty applied to any action whose total delay exceeds the limit is assumed to teach the policy to satisfy all per-task deadlines and node-capacity constraints in problem (2), but the paper never measures whether it actually does.

Editorial extensions

If this is right

  • If the reported margins hold, SARMTO would lower operating cost for IoT task offloading across heterogeneous MEC and cloud nodes by up to 40% compared with state-of-the-art DRL baselines.
  • The 41.5% energy-efficiency improvement over CO at large data sizes implies substantial savings in battery-constrained IoT deployments, especially when tasks carry large data payloads.
  • The framework's adaptive security overhead makes it possible to co-optimize encryption and hashing cost with delay and energy, so security can be provided without a fixed performance tax.
  • Because AC-DQN embeds constraints in the Q-function, the approach claims to avoid the convergence issues of penalty-based rewards, potentially extending to other hard-constrained scheduling problems beyond serverless edge computing.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper, the real test of the action-constraint idea is constraint satisfaction: a natural follow-up is to report the rate of deadline and capacity violations under the trained policy, not just average cost and delay.
  • Beyond the paper, the mechanism's generality suggests it could be applied to other combinatorial allocation problems with hard feasibility constraints, where a fixed penalty in the value function may be cheaper than full constrained policy optimization.
  • Beyond the paper, the paper leaves the penalty magnitude $\lambda$ fixed and untuned; a minimal extension would sweep $\lambda$ and measure whether larger penalties actually drive violation rates to zero, or whether they merely increase Q-value variance.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes SARMTO, a framework for secure resource allocation in serverless multi-cloud edge computing, built around AC-DQN, a deep Q-network with a claimed 'action constraint' mechanism. The authors formulate the problem as a constrained optimization over task offloading decisions, model delay and energy costs with an added security overhead, and evaluate SARMTO against five baselines in a custom Python simulator. The paper reports substantial cost and energy reductions, up to 40% and 41.5% respectively over baselines. The central claims are that AC-DQN respects system constraints, that its action-constraint mechanism is preferable to penalty-based DRL, and that the adaptive security mechanism balances security with overhead. As presented, however, the action-constraint mechanism is itself a penalty on aggregate delay, the reported experiments lack statistical grounding, and the security contribution is not evaluated as security. These issues undermine the paper's main conclusions.

Significance. If the results held as stated, a 20-40% system-cost reduction in a realistic serverless multi-cloud edge setting would be practically valuable, and an action-constrained DRL method that genuinely enforces per-task deadlines and node capacities would be a useful contribution to constrained DRL. The paper has a clear MDP formulation and a readable system model, and it does include five baselines. However, the central algorithmic innovation is not actually an action constraint: Eq. (7) is a fixed penalty on total delay, and the problem's per-task and capacity constraints are never enforced or even measured. The experiments report single-point numbers without error bars, multiple seeds, or significance tests, and the security mechanism is modeled only as added CPU cycles. The significance of the claimed improvements cannot be assessed from the evidence provided, and the paper does not substantiate its 'secure' and 'constraint-satisfying' contributions.

major comments (4)
  1. [§4.3, Eq. (7)-(8), Algorithm 2] The claim that AC-DQN is 'action-constrained' and avoids the drawbacks of penalty-based methods is not supported. The function f_constraint in Eq. (7) returns a constant -λ when the total delay T_total(a_j) exceeds T_max; it is a scalar penalty on aggregate delay, not a restriction of the action space, and it does not enforce the per-task deadlines T_ij ≤ T_ij or the node-capacity constraints Σ x_ijk C_ij ≤ f_k of problem (2). T_max is not defined in problem (2), which uses per-task deadlines. Moreover, the penalty appears only in the Q-target (Eq. (8), Algorithm 2 line 17), not in the immediate reward r_t, and exploration (Algorithm 2 line 9) samples actions uniformly with no feasibility filtering. No bound or tuning of λ is provided. Since Section 5 reports no deadline-miss ratio or capacity-violation metric, the reported system-cost reductions may reflect infeasible schedules, and the alleged superiority over penalty-based baselines is not established.
  2. [§5.1-5.5] The experimental evaluation is not statistically grounded. All reported results are single-point estimates, e.g., '23.6% reduction' at 1000 tasks and '41.5% improvement' in energy efficiency, with no error bars, no multiple seeds, and no significance tests. Since DQN training is stochastic, this is a serious omission. The baselines are also weak: local computation, random offloading, and two DQN variants, with no comparison to recent constrained DRL or serverless scheduling methods from the related work, such as [22] or [30]. The claimed improvements cannot be separated from run-to-run variance, and the headline numbers in the abstract are therefore not supported.
  3. [§4.4, Eqs. (9)-(11)] The 'adaptive security mechanism' is modeled only as additional CPU cycles for encryption, decryption, and hashing, and these overheads are added directly to the cost function. There is no threat model, no metric of security effectiveness (e.g., attack success rate, confidentiality, integrity verification failures), and no experiment that varies the threat landscape or task sensitivity. The paper claims 'robust security without compromising efficiency,' but the only security-related quantities measured are time and energy overheads. The security contribution is therefore not evaluated, and the term 'adaptive' is not operationalized in the experiments.
  4. [§5.1-5.5, Eqs. (3)-(4), (11)] The evaluation is entirely in-simulation, and the simulator computes system cost from the same delay and energy models that define the agent's reward function. This means the reported cost and energy numbers are in-model measures rather than independent benchmarks. No real trace data, validated channel model, or comparison to physical measurements is provided. The paper should explicitly acknowledge that the results only compare policies under the authors' own cost model, and should temper the abstract's claims of 'state-of-the-art' improvement accordingly.
minor comments (5)
  1. [Conclusion] The reference list contains an incomplete citation placeholder '[?,44–54]' in the concluding paragraph; this must be resolved before any resubmission.
  2. [§5.5] The text says 'As illustrated in Figure 4(a)' when discussing the impact of MEC computational capacity, but Figure 4 shows results for data size; the relevant figure appears to be Figure 2(a)-(d). The figure references should be corrected.
  3. [§4.3, §5.1] Several hyperparameters are either not reported or only given as examples: the penalty λ is said to be 'e.g., 1000' with no sensitivity analysis, and the exploration schedule ϵ_0, ϵ_decay, ϵ_min, network sizes, replay buffer capacity, and minibatch size are not collected in one place. A parameter table would improve reproducibility.
  4. [§4.5, Algorithm 2] The adaptive exploration strategy of Eq. (13) is described in the text, but Algorithm 2 uses a constant ϵ in line 6 and does not show how ϵ(t) is computed or updated. The algorithm and the formal definition should be consistent.
  5. [§4.4] The security protocol uses MD5 for hashing; MD5 is known to be vulnerable to collision attacks, so its use for data integrity is questionable and should be justified or replaced with a collision-resistant hash in a paper claiming 'robust security.'

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: reported numbers are simulator outputs rather than fitted or self-cited predictions; the main caveat (Eq. 7 is a penalty, not a hard constraint) affects validity and novelty, not circularity.

full rationale

The paper's derivation chain does not contain a prediction that is equivalent to its inputs by construction. The MDP reward is defined as the negative of the system cost, and the paper then evaluates system cost; this is objective alignment inherent to RL, not a fitted-input cycle. All reported cost and energy figures come from the authors' own simulator across task counts, data sizes, and MEC capacities, and no parameter is fitted to a target dataset and then renamed as a prediction. The table of CPU cycles per byte is cited from prior published work and merely parameterizes the simulation rather than assuming the conclusion. The most serious concern is in Sec. 4.3: Eq. (7) defines fconstraint as a fixed penalty added to target Q-values, so the claimed contrast between 'action-constrained' learning and 'penalty-based' DRL is not substantiated, and the constraint mechanism does not by itself enforce the per-task and capacity constraints of problem (2). However, that is a correctness and novelty risk, not a circularity: the comparative performance results are empirical simulation outputs rather than consequences of Eq. (7) by definition. Therefore the circularity score is low.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central results rest on hand-chosen weights and penalty constants, on an unverified constraint-enforcement assumption, and on a simplified simulator whose cost model is authored by the same team. No new physical entities are introduced.

free parameters (5)
  • Penalty constant lambda = 1000
    Set by hand in Eq (7); controls constraint enforcement, never tuned or justified.
  • Cost weights alpha1 and alpha2 = 0.5, 0.5
    Chosen in §5.1 without sensitivity analysis; they define the objective and therefore all cost comparisons.
  • Exploration parameters epsilon0, epsilon_decay, epsilon_min = not reported
    Eq (13) introduces these three parameters but the paper gives no values.
  • Q-network hyperparameters = not reported
    Hidden sizes, learning rate, batch size, replay capacity, target update rate are not specified in §4.3 or §5.1.
  • System simulation parameters = 20 MHz bandwidth, loss factor 4, noise 1e-13, 10 GHz/100 GHz capacities, 1 km/10 km distances
    Chosen in §5.1 without sensitivity analysis; all comparative results live inside this simplified environment.
assumptions (4)
  • domain assumption Task complexity mapping Cij = zeta(Dij) (Table 1) is accurate for real applications
    Eq (1) and Table 1 assume fixed CPU-cycles-per-byte per application; errors here change every delay and energy result.
  • domain assumption The security overhead model captures real cryptographic costs and provides meaningful security
    §4.4 uses RSA, MD5, and fixed CPU-cycle numbers; MD5 is cryptographically broken and no threat model or adaptive adjustment is defined.
  • ad hoc to paper A single scalar penalty on total delay suffices to enforce all per-task and node-capacity constraints
    Eq (7) checks only total delay against Tmax, whereas problem (2) has per-task deadlines and capacity constraints; no proof or verification is offered.
  • domain assumption The deep Q-network converges to a stable policy in this MDP
    Training is asserted to work over 1000 episodes; no convergence analysis or learning curves across seeds are provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Secure Resource Allocation via Constrained Deep Reinforcement Learning." pith.science (2026). https://pith.science/paper/RYPI6PGJ

@misc{pith2026250111557,
  author       = {Pith},
  title        = {Pith review of: Secure Resource Allocation via Constrained Deep Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RYPI6PGJ}},
  note         = {Machine review of arXiv:2501.11557}
}
read the original abstract

The proliferation of Internet of Things (IoT) devices and the advent of 6G technologies have introduced computationally intensive tasks that often surpass the processing capabilities of user devices. Efficient and secure resource allocation in serverless multi-cloud edge computing environments is essential for supporting these demands and advancing distributed computing. However, existing solutions frequently struggle with the complexity of multi-cloud infrastructures, robust security integration, and effective application of traditional deep reinforcement learning (DRL) techniques under system constraints. To address these challenges, we present SARMTO, a novel framework that integrates an action-constrained DRL model. SARMTO dynamically balances resource allocation, task offloading, security, and performance by utilizing a Markov decision process formulation, an adaptive security mechanism, and sophisticated optimization techniques. Extensive simulations across varying scenarios, including different task loads, data sizes, and MEC capacities, show that SARMTO consistently outperforms five baseline approaches, achieving up to a 40% reduction in system costs and a 41.5% improvement in energy efficiency over state-of-the-art methods. These enhancements highlight SARMTO's potential to revolutionize resource management in intricate distributed computing environments, opening the door to more efficient and secure IoT and edge computing applications.

Figures

Figures reproduced from arXiv: 2501.11557 by the authors.

Figure 1
Figure 1. Serverless multi-cloud edge computing model [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Performance under different MEC computational capacities, including [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Performance under different number of tasks, including system cost (a) [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Performance under different data sizes, including system cost (a) and [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 38 canonical work pages

  1. [22]

    Distributed task scheduling in serverless edge computing networks for the internet of things: A learning approach,

    Q. Tang, R. Xie, F. R. Yu, T. Chen, R. Zhang, T. Huang, and Y. Liu, “Distributed task scheduling in serverless edge computing networks for the internet of things: A learning approach,” IEEE Internet Things J. , vol. 9, no. 20, pp. 19634–19648, 2022

    work page 2022

  2. [30]

    Security computing resource allocation based on deep reinforcement learning in serverless multi-cloud edge computing,

    H. Zhang, J. Wang, H. Zhang, and C. Bu, “Security computing resource allocation based on deep reinforcement learning in serverless multi-cloud edge computing,” Future Generation Computer Systems , vol. 151, pp. 152–161, 2024

    work page 2024

  3. [1]

    Efficient par- allel split learning over resource-constrained wireless edge networks,

    Z. Lin, G. Zhu, Y. Deng, X. Chen, Y. Gao, K. Huang, and Y. Fang, “Efficient par- allel split learning over resource-constrained wireless edge networks,”IEEE Trans- actions on Mobile Computing , 2024

    work page 2024

  4. [2]

    Constructing 4d radio map in leo satellite networks with limited samples,

    H. Yuan, Z. Chen, Z. Lin, J. Peng, Y. Zhong, X. Hu, S. Xue, W. Li, and Y. Gao, “Constructing 4d radio map in leo satellite networks with limited samples,”arXiv preprint arXiv:2501.02775, 2025

    arXiv 2025

  5. [3]

    Leo- split: A semi-supervised split learning framework over leo satellite networks,

    Z. Lin, Y. Zhang, Z. Chen, Z. Fang, C. Wu, X. Chen, Y. Gao, and J. Luo, “Leo- split: A semi-supervised split learning framework over leo satellite networks,”arXiv preprint arXiv:2501.01293, 2025

    arXiv 2025

  6. [4]

    Vulseye: Detect Smart Contract Vulnerabilities via Stateful Directed Graybox Fuzzing

    R. Liang, J. Chen, C. Wu, K. He, Y. Wu, R. Cao, R. Du, Y. Liu, and Z. Zhao, “Vuls- eye: Detect smart contract vulnerabilities via stateful directed graybox fuzzing,” arXiv preprint arXiv:2408.10116 , 2024

    work page Pith review arXiv 2024

  7. [5]

    Graphlearningformulti-satellitebasedspectrumsensing,

    H. Yuan, Z. Chen, Z. Lin, J. Peng, Z. Fang, Y. Zhong, Z. Song, X. Wang, and Y.Gao,“Graphlearningformulti-satellitebasedspectrumsensing,” in Proc. ICCT, 2023, pp. 1112–1116

    work page 2023

  8. [6]

    Optimizing the learning performance in mobile augmented reality systems with cnn,

    Y. He, J. Ren, G. Yu, and Y. Cai, “Optimizing the learning performance in mobile augmented reality systems with cnn,”IEEE Transactions on Wireless Communi- cations, vol. 19, no. 8, pp. 5333–5344, 2020

    work page 2020

Show all 54 references
  1. [7]

    Deep-learning-incorporated aug- mented reality application for engineering lab training,

    J. Estrada, S. Paheding, X. Yang, and Q. Niyaz, “Deep-learning-incorporated aug- mented reality application for engineering lab training,”Applied Sciences, vol. 12, no. 10, p. 5159, 2022

    2022

  2. [8]

    Integrated registration and oc- clusion handling based on deep learning for augmented-reality-assisted assembly instruction,

    W. Li, J. Wang, M. Liu, S. Zhao, and X. Ding, “Integrated registration and oc- clusion handling based on deep learning for augmented-reality-assisted assembly instruction,” IEEE Transactions on Industrial Informatics, vol. 19, no. 5, pp. 6825– 6835, 2022

    2022

  3. [9]

    Tracking and transmission design in terahertz v2i networks,

    Z. Lin, L. Wang, J. Ding, Y. Xu, and B. Tan, “Tracking and transmission design in terahertz v2i networks,”IEEE Transactions on Wireless Communications , vol. 22, no. 6, pp. 3586–3598, 2022

    2022

  4. [10]

    Ic3m: In-car multi- modal multi-object monitoring for abnormal status of both driver and passengers,

    Z. Fang, Z. Lin, S. Hu, H. Cao, Y. Deng, X. Chen, and Y. Fang, “Ic3m: In-car multi- modal multi-object monitoring for abnormal status of both driver and passengers,” arXiv preprint arXiv:2410.02592 , 2024

    2024 arXiv

  5. [11]

    Channel power gain estimation for terahertz vehicle-to-infrastructure networks,

    Z. Lin, L. Wang, J. Ding, B. Tan, and S. Jin, “Channel power gain estimation for terahertz vehicle-to-infrastructure networks,” IEEE Communications Letters , vol. 27, no. 1, pp. 155–159, 2022

    2022

  6. [12]

    Rethinking membership inference attacks against transfer learning,

    C. Wu, J. Chen, Q. Fang, K. He, Z. Zhao, H. Ren, G. Xu, Y. Liu, and Y. Xi- ang, “Rethinking membership inference attacks against transfer learning,”IEEE Transactions on Information Forensics and Security , 2024

    2024

  7. [13]

    Fedac: A adaptive clus- tered federated learning framework for heterogeneous data,

    Y. Zhang, H. Chen, Z. Lin, Z. Chen, and J. Zhao, “Fedac: A adaptive clus- tered federated learning framework for heterogeneous data,” arXiv preprint arXiv:2403.16460, 2024

    2024 arXiv

  8. [14]

    Multi-layer computation offload- ing in distributed heterogeneous mobile edge computing networks,

    P. Wang, B. Di, L. Song, and N. R. Jennings, “Multi-layer computation offload- ing in distributed heterogeneous mobile edge computing networks,”IEEE Trans. Cognitive Commun. Networking , vol. 8, no. 2, pp. 1301–1315, 2022

    2022

  9. [15]

    Secure channel establishment scheme for task delivery in vehicular cloud computing,

    R. Zhang, L. Zhang, Q. Wu, and J. Zhou, “Secure channel establishment scheme for task delivery in vehicular cloud computing,”IEEE Transactions on Information Forensics and Security, 2024

    2024

  10. [16]

    Hierarchical split federated learning: Convergence analysis and system optimization,

    Z. Lin, W. Wei, Z. Chen, C.-T. Lam, X. Chen, Y. Gao, and J. Luo, “Hierarchical split federated learning: Convergence analysis and system optimization,” arXiv preprint arXiv:2412.07197, 2024

    2024 arXiv

  11. [17]

    Adaptive resource allo- cation for semantic communication networks,

    L. Wang, W. Wu, F. Zhou, Z. Yang, Z. Qin, and Q. Wu, “Adaptive resource allo- cation for semantic communication networks,”IEEE Transactions on Communi- cations, 2024

    2024

  12. [18]

    Resource allocation and trust computing for blockchain-enabled edge computing system,

    L. Zhang, Y. Zou, W. Wang, Z. Jin, Y. Su, and H. Chen, “Resource allocation and trust computing for blockchain-enabled edge computing system,”Computers & Security, vol. 105, p. 102249, 2021

    2021

  13. [19]

    Resourceallocation and workload scheduling for large-scale distributed deep learning: A survey,

    F.Liang,Z.Zhang,H.Lu,C.Li,V.Leung,Y.Guo,andX.Hu,“Resourceallocation and workload scheduling for large-scale distributed deep learning: A survey,”arXiv preprint arXiv:2406.08115, 2024

    2024 arXiv

  14. [20]

    Fedsn: A federated learning framework over heterogeneous leo satellite networks,

    Z. Lin, Z. Chen, Z. Fang, X. Chen, X. Wang, and Y. Gao, “Fedsn: A federated learning framework over heterogeneous leo satellite networks,”IEEE Transactions on Mobile Computing , 2024

    2024

  15. [21]

    Privacyeafl: Privacy-enhanced ag- gregation for federated learning in mobile crowdsensing,

    M. Zhang, S. Chen, J. Shen, and W. Susilo, “Privacyeafl: Privacy-enhanced ag- gregation for federated learning in mobile crowdsensing,”IEEE Transactions on Information Forensics and Security , 2023

    2023

  16. [23]

    Performance optimization of serverless edge computing function offloading based on deep reinforcement learning,

    X. Yao, N. Chen, X. Yuan, and P. Ou, “Performance optimization of serverless edge computing function offloading based on deep reinforcement learning,”Future Generation Computer Systems , vol. 139, pp. 74–86, 2023

    2023

  17. [24]

    Learning-based privacy-aware offloading for healthcare iot with energy harvesting,

    M. Min, X. Wan, L. Xiao, Y. Chen, M. Xia, D. Wu, and H. Dai, “Learning-based privacy-aware offloading for healthcare iot with energy harvesting,”IEEE Internet Things J., vol. 6, no. 3, pp. 4307–4316, 2018

    2018

  18. [25]

    Security modeling and efficient computation offloading for service workflow in mobile edge computing,

    B. Huang, Z. Li, P. Tang, S. Wang, J. Zhao, H. Hu, W. Li, and V. Chang, “Security modeling and efficient computation offloading for service workflow in mobile edge computing,” Future Generation Computer Systems , vol. 97, pp. 755–774, 2019

    2019

  19. [26]

    Performance optimization of serverless comput- ing for latency-guaranteed and energy-efficient task offloading in energy-harvesting industrial iot,

    H. Ko, S. Pack, and V. C. Leung, “Performance optimization of serverless comput- ing for latency-guaranteed and energy-efficient task offloading in energy-harvesting industrial iot,” IEEE Internet Things J. , vol. 10, no. 3, pp. 1897–1907, 2021

    1907

  20. [27]

    Architecture and performance evalu- ation of distributed computation offloading in edge computing,

    C. Cicconetti, M. Conti, and A. Passarella, “Architecture and performance evalu- ation of distributed computation offloading in edge computing,”Simulation Mod- elling Practice and Theory , vol. 101, p. 102007, 2020

    2020

  21. [28]

    Multi-cloud provisioning and load distribution for three- tier applications,

    N. Grozev and R. Buyya, “Multi-cloud provisioning and load distribution for three- tier applications,” ACM Trans. Autonomous Adaptive Systems , vol. 9, no. 3, pp. 1–21, 2014

    2014

  22. [29]

    Automated feder- ated pipeline for parameter-efficient fine-tuning of large language models,

    Z. Fang, Z. Lin, Z. Chen, X. Chen, Y. Gao, and Y. Fang, “Automated feder- ated pipeline for parameter-efficient fine-tuning of large language models,”arXiv preprint arXiv:2404.06448, 2024

    2024 arXiv

  23. [31]

    Efficientand secure multi-user multi-task computation offloading for mobile-edge computing in mobile iot networks,

    I.A.Elgendy,W.-Z.Zhang,Y.Zeng,H.He,Y.-C.Tian,andY.Yang,“Efficientand secure multi-user multi-task computation offloading for mobile-edge computing in mobile iot networks,” IEEE Trans. Network Service Management , vol. 17, no. 4, pp. 2410–2422, 2020

    2020

  24. [32]

    Service offloading with deep q-network for digital twinning-empowered internet of vehicles in edge computing,

    X. Xu, B. Shen, S. Ding, G. Srivastava, M. Bilal, M. R. Khosravi, V. G. Menon, M. A. Jan, and M. Wang, “Service offloading with deep q-network for digital twinning-empowered internet of vehicles in edge computing,”IEEE Trans. Indus- trial Informatics, vol. 18, no. 2, pp. 1414–...

    2020

  25. [33]

    Multiuser computation offloading and resource allocation for cloud–edge heterogeneous network,

    Q. Chen, Z. Kuang, and L. Zhao, “Multiuser computation offloading and resource allocation for cloud–edge heterogeneous network,”IEEE Internet Things J. , vol. 9, no. 5, pp. 3799–3811, 2021

    2021

  26. [34]

    Echohand: High accu- racy and presentation attack resistant hand authentication on commodity mobile devices,

    C. Wu, J. Chen, K. He, Z. Zhao, R. Du, and C. Zhang, “Echohand: High accu- racy and presentation attack resistant hand authentication on commodity mobile devices,” in Proceedings of the 2022 ACM SIGSAC conference on computer and communications security, 2022

    2022

  27. [35]

    Rethinking adversarial attacks in reinforcement learning from policy distribution perspective,

    T.Duan,Z.Zhang,Z.Lin,Y.Gao,L.Xiong,Y.Cui,H.Liang,X.Chen,H.Cui,and D. Huang, “Rethinking adversarial attacks in reinforcement learning from policy distribution perspective,” arXiv preprint arXiv:2501.03562 , 2025

    2025 arXiv

  28. [36]

    Toward robust detection of pup- pet attacks via characterizing fingertip-touch behaviors,

    C. Wu, K. He, J. Chen, Z. Zhao, and R. Du, “Toward robust detection of pup- pet attacks via characterizing fingertip-touch behaviors,” IEEE Transactions on Dependable and Secure Computing , vol. 19, no. 6, pp. 4002–4018, 2021

    2021

  29. [37]

    Sok: Comprehensive security overview, challenges, and future directions of voice-controlled systems,

    H. Xu, C. Wu, Y. Gu, X. Shang, J. Chen, K. He, and R. Du, “Sok: Comprehensive security overview, challenges, and future directions of voice-controlled systems,” arXiv preprint arXiv:2405.17100 , 2024

    2024 arXiv

  30. [38]

    Liveness is not enough: Enhancing fingerprint authentication with behavioral biometrics to defeat puppet attacks,

    C. Wu, K. He, J. Chen, Z. Zhao, and R. Du, “Liveness is not enough: Enhancing fingerprint authentication with behavioral biometrics to defeat puppet attacks,” in 29th USENIX Security Symposium (USENIX Security 20) , 2020, pp. 2219–2236

    2020

  31. [39]

    Pushing large lan- guage models to the 6g edge: Vision, challenges, and opportunities,

    Z. Lin, G. Qu, Q. Chen, X. Chen, Z. Chen, and K. Huang, “Pushing large lan- guage models to the 6g edge: Vision, challenges, and opportunities,”arXiv preprint arXiv:2309.16739, 2023

    2023 arXiv

  32. [40]

    Semantic sleuth: Identifying ponzi contracts via large language models,

    C. Wu, J. Chen, Z. Wang, R. Liang, and R. Du, “Semantic sleuth: Identifying ponzi contracts via large language models,” inProceedings of the 39th IEEE/ACM International Conference on Automated Software Engineering , 2024, pp. 582–593

    2024

  33. [41]

    Adaptsfl: Adaptive split federated learning in resource-constrained edge networks,

    Z. Lin, G. Qu, W. Wei, X. Chen, and K. K. Leung, “Adaptsfl: Adaptive split federated learning in resource-constrained edge networks,” arXiv preprint arXiv:2403.13101, 2024

    2024 arXiv

  34. [42]

    Lcfed: An efficient clus- tered federated learning framework for heterogeneous data,

    Y. Zhang, H. Chen, Z. Lin, Z. Chen, and J. Zhao, “Lcfed: An efficient clus- tered federated learning framework for heterogeneous data,” arXiv preprint arXiv:2501.01850, 2025

    2025 arXiv

  35. [43]

    Split learning in 6g edge networks,

    Z. Lin, G. Qu, X. Chen, and K. Huang, “Split learning in 6g edge networks,”IEEE Wireless Communications, 2024

    2024

  36. [44]

    Source code summarization in the era of large language models,

    W. Sun, Y. Miao, Y. Li, H. Zhang, C. Fang, Y. Liu, G. Deng, Y. Liu, and Z. Chen, “Source code summarization in the era of large language models,” CoRR, vol. abs/2407.07959, no. 1, pp. 1–13, 2024

    2024 arXiv

  37. [45]

    A survey of source code search: A 3-dimensional perspective,

    W. Sun, C. Fang, Y. Ge, Y. Hu, Y. Chen, Q. Zhang, X. Ge, Y. Liu, and Z. Chen, “A survey of source code search: A 3-dimensional perspective,”ACM Transactions on Software Engineering and Methodology , vol. 33, no. 6, pp. 166:1–51, 2024

    2024

  38. [46]

    An extractive-and-abstractive framework for source code summarization,

    W. Sun, C. Fang, Y. Chen, Q. Zhang, G. Tao, Y. You, T. Han, Y. Ge, Y. Hu, B. Luo, and Z. Chen, “An extractive-and-abstractive framework for source code summarization,” ACM Transactions on Software Engineering and Methodology , vol. 33, no. 3, pp. 75:1–75:39, 2024

    2024

  39. [47]

    Esale: Enhancing code-summary alignment learning for source code summarization,

    C. Fang, W. Sun, Y. Chen, X. Chen, Z. Wei, Q. Zhang, Y. You, B. Luo, Y. Liu, and Z. Chen, “Esale: Enhancing code-summary alignment learning for source code summarization,” IEEE Transactions on Software Engineering (Early Access) , pp. 1–18, 2024

    2024

  40. [48]

    Abstract syntax tree for programming language understand- ing and representation: How far are we?

    W. Sun, C. Fang, Y. Miao, Y. You, M. Yuan, Y. Chen, Q. Zhang, A. Guo, X. Chen, Y. Liu, and Z. Chen, “Abstract syntax tree for programming language understand- ing and representation: How far are we?”CoRR, vol. abs/2312.00413, no. 1, pp. 1–47, 2023

    2023 arXiv

  41. [49]

    Maf: Method-anchored test fragmentation for test code plagiarism detection,

    W. Sun, X. Wang, H. Wu, D. Duan, Z. Sun, and Z. Chen, “Maf: Method-anchored test fragmentation for test code plagiarism detection,” inProceedings of the 41th International Conference on Software Engineering, Software Engineering Educa- tion and Training. Montreal, QC, Canada: ...

    2019

  42. [50]

    Backdooring neural code search,

    W. Sun, Y. Chen, G. Tao, C. Fang, X. Zhang, Q. Zhang, and B. Luo, “Backdooring neural code search,” inProceedings of the 61st Annual Meeting of the Association for Computational Linguistics . Toronto, Canada: Association for Computational Linguistics, July 9-14 2023, pp. 9692–9708

    2023

  43. [51]

    Eliminating backdoors in neural code models via trigger inversion,

    W. Sun, Y. Chen, C. Fang, Y. Feng, Y. Xiao, A. Guo, Q. Zhang, Y. Liu, B. Xu, and Z. Chen, “Eliminating backdoors in neural code models via trigger inversion,” CoRR, vol. abs/2408.04683, no. 1, pp. 1–12, 2024

    2024 arXiv

  44. [52]

    Mu- tual information guided backdoor mitigation for pre-trained encoders,

    T. Han, W. Sun, Z. Ding, C. Fang, H. Qian, J. Li, Z. Chen, and X. Zhang, “Mu- tual information guided backdoor mitigation for pre-trained encoders,”CoRR, vol. abs/2406.03508, no. 1, pp. 1–12, 2024

    2024

  45. [53]

    On the effectiveness of distillation in mitigating backdoors in pre-trained encoder,

    T. Han, S. Huang, Z. Ding, W. Sun, Y. Feng, C. Fang, J. Li, H. Qian, C. Wu, Q. Zhang, Y. Liu, and Z. Chen, “On the effectiveness of distillation in mitigating backdoors in pre-trained encoder,” CoRR, vol. abs/2403.03846, no. 1, pp. 1–17, 2024

    2024

  46. [54]

    RULER: discriminative and iterative adversarial training for deep neural network fairness,

    G. Tao, W. Sun, T. Han, C. Fang, and X. Zhang, “RULER: discriminative and iterative adversarial training for deep neural network fairness,” inProceedings of the 30th Joint European Software Engineering Conference and Symposium on the Foundations of Software Engineering . Singa...

    2022

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.