Mustafa Noori Shukur

Mr. Mustafa Noori Shukur

orcid

Research Focus & Metrics

Academic Career Length

1Year (2026-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
13 August 2026 in Preprints.org
Verified
Preprint
Open access
PDF file
NSNoor Saeed Saleh
Noor Saeed Saleh
Maath Frman IsmaeelMaath Frman Ismaeel
Maath Frman Ismaeel
MSMustafa Noori Shukur
Mustafa Noori Shukur
KMKholood Jamal Moulood
Kholood Jamal Moulood
HAHachel Amer Ayash
Hachel Amer Ayash
Qutaiba AlasadQutaiba Alasad
Qutaiba Alasad
Autonomous multi-agent reinforcement learning (MARL) has been increasingly entrusted with real-time control of smart energy infrastructures, where agents remain vulnerable to the false data injection attacks (FDIA), Byzantine failures, and specification misalignment. Existing proposed constitutional AI techniques target discrete language-based coordination and offer no protection for continuous-action cyber-physical control. In this paper, we propose SafeAgentGrid, a hierarchical neuro-symbolic constitutional framework integrating (1) constitutional agents whose actions pass a projection filter that can guarantee every hard box-type constraint by the construction; (2) a hybrid neural-surrogate mediator whose low-latency fast path has been admitted only after a split-conformal test bounding its alignment shortfall, and whose full path resolves a constrained program over a tightened DC reachable set, fairness (Gini) and risk-aware security weighting; and (3) 8-bit quantized models for edge computing. On a DC power-flow simulation with a ring-topology susceptance matrix (10 agents, 40 evaluation episodes), the proposed SafeAgentGrid accomplishes 200,186.9 average reward, 0.0125 ± 0.0014 violations/step (0.0194 ± 0.0016 under 20% FDIA; 0.0140 ± 0.0015 under the Byzantine failure of \( \lfloor N/3\rfloor \) agents), 100% grid efficiency, 0.999 alignment and 0.92 ms/step latency, outperforming four strong related techniques, trained in the same environment, in both reward and safety (Wilcoxon signed-rank, p ≤ 3.4 × 10-8). We further show that tightening the reachable set is really necessary for the safety guarantee in order to survive stochastic realization, that robust repair must be anchored on a per-agent physical reference rather than the population median, and that the effectiveness of the security module is governed by the weight with which the mediator delegates authority to the agents.

5 June 2026 in Preprints.org
Verified
Preprint
Open access
PDF file
Maath Frman IsmaeelMaath Frman Ismaeel
Maath Frman Ismaeel
KMKholood Jamal Moulood
Kholood Jamal Moulood
MSMustafa Noori Shukur
Mustafa Noori Shukur
EHEkram H. Hasan
Ekram H. Hasan
OAOqbah Salim Atiyah
Oqbah Salim Atiyah
Qutaiba AlasadQutaiba Alasad
Qutaiba Alasad
The rapid advancement of modern Deep Neural Networks (DNNs) has played a crucial role in aiding humans in many real-world applications, yet their hardware accelerators have been proven to be vulnerable to malicious attacks. One particularly severe and serious attack involves inserting a hardware Trojan (HT) into DNN accelerator hardware in order to enable attackers to stealthily manipulate model predictions during the supply chain. In this paper, we present a stealthy HT that is difficult to be detected and has a significant impact on the performance of DNN models. To successfully achieve this goal, we introduce the Sensitivity-Based Weight Selection (SBWS) algorithm, a novel technique that adapts machine learning (ML) sensitivity analysis to identify and modify only a small number of weights that have the highest impact on DNN performance, compared to previous work. We evaluate our proposed attack on five DNN models and multiple datasets using two designed payload types: weight zeroing and sign-flipping, and record the results based on various security metrics. The experimental results show average accuracy reductions of 26.7% for the zeroing attack and 48.1% for the sign-flipping attack, yielding an overall average of 37.4%, calculated over five independent runs per dataset with standard deviation < 2%. The sign-flipping technique consistently outperforms zeroing one because it preserves the magnitudes of the attacked weights while inverting their signs, thereby disrupting the learned decision boundaries more severely and amplifying error propagation in subsequent layers. These results significantly exceed previous random-weight perturbation attacks (typically 12–20% drops) and other targeted HT approaches, while incurring lower computational and hardware resource overheads. This work provides a more effective and scalable method for assessing the vulnerability of DNN accelerators under real supply-chain threat models.

Verified
Journal Article
Open access
PDF file
Maath Frman IsmaeelMaath Frman Ismaeel
Maath Frman Ismaeel
KMKholood Jamal Moulood
Kholood Jamal Moulood
MSMustafa Noori Shukur
Mustafa Noori Shukur
EHEkram H. Hasan
Ekram H. Hasan
OAOqbah Salim Atiyah
Oqbah Salim Atiyah
Qutaiba AlasadQutaiba Alasad
Qutaiba Alasad
The rapid advancement of modern deep neural networks (DNNs) has played a crucial role in aiding humans across many real-world applications; yet, their hardware accelerators have been proven to be vulnerable to malicious attacks. One particularly severe and serious attack involves inserting a hardware Trojan (HT) into DNN accelerator hardware in order to enable attackers to stealthily manipulate model predictions during the supply chain. In this paper, we present a possible stealthy HT architecture that is difficult to detect and has a significant impact on the performance of DNN models. To successfully achieve this goal, we introduce the Sensitivity-Based Weight Selection (SBWS) algorithm, a novel technique that adapts machine learning (ML) sensitivity analysis to identify and modify a small number of weights that have the highest impact on DNN performance, compared to previous work. We evaluate the proposed attack on five DNN model tests (two distinct DNN models and four different datasets) using two designed payload types (weight zeroing and sign-flipping) and record the results based on various security metrics. The experimental results show average accuracy reductions of 26.7% for the zeroing attack and 48.1% for the sign-flipping attack, yielding an overall average of 37.4%, calculated over five independent runs per dataset with standard deviation <2%. The sign-flipping technique consistently outperforms zeroing because it preserves the magnitudes of the attacked weights while inverting their signs, thereby disrupting the learned decision boundaries more severely and amplifying error propagation in subsequent layers. These results significantly exceed those of previous random-weight perturbation attacks (typically 12–20% drops) and other targeted HT approaches while incurring lower computational and hardware resource overheads. This work provides a more effective and scalable method for assessing the vulnerability of DNN accelerators under real supply chain threat models.

10 October 2026 in Network
Verified
Journal Article
Open access
PDF file
NSNoor Saeed Saleh
Noor Saeed Saleh
Maath Frman IsmaeelMaath Frman Ismaeel
Maath Frman Ismaeel
MSMustafa Noori Shukur
Mustafa Noori Shukur
KMKholood Jamal Moulood
Kholood Jamal Moulood
HAHachel Amer Ayash
Hachel Amer Ayash
Qutaiba AlasadQutaiba Alasad
Qutaiba Alasad
Multi-agent reinforcement learning (MARL) has been increasingly utilized for the real-time control of smart energy infrastructures, where agents are still vulnerable to false data injection attacks (FDIAs), Byzantine failures, and specification misalignment. Existing constitutional AI techniques target discrete language-based coordination, not continuous-action cyber–physical control. In this paper, we propose SafeAgentGrid, a hierarchical neuro-symbolic constitutional framework. A projection filter guarantees every hard box-type constraint by construction. A hybrid neural-surrogate mediator combines a low-latency fast path, admitted only after a split-conformal test bounds its alignment shortfall, with a full path over a tightenedDC reachable set under fairness (Gini) and risk-aware security weighting. Finally, 8-bit quantized models support edge computing. In a data-informed DC power-flow simulation, SafeAgentGrid accomplishes an average reward of 200,186.9 at 0.0125±0.0014 violations/step, rising only to 0.0194±0.0016 under 20% FDIA and 0.0140±0.0015 under Byzantine failure of ⌊N/3⌋ agents, at 100% grid efficiency and 0.79 ms/step mediation latency. It enhances both reward and safety over four unmediated techniques trained in the same environment (Wilcoxon signed-rank, p≤3.4×10−8) on all ten training seeds and outperforms a joint safety shield and a Lagrangian constrained MARL technique, which violate the limits on 57.5% and 98.8% of the steps. We further derive the tightening from a chance constraint and show that it is necessary under the stochastic realization.