Ahmad Ghasemi

Department of Electrical & Computer Engineering

University of Massachusetts Amherst

aghasemi (at) umass [dot] edu

I am a Postdoctoral Research Fellow and Adjunct Faculty in Electrical & Computer Engineering at the University of Massachusetts Amherst. My research builds dependable Edge AI systems for networked, resource-constrained platforms (IoT/embedded/CPS and multi-agent autonomy), where decisions must be made under hard constraints on latency, energy, and memory, and where dependence, partial observability, and distribution shift are the default operating conditions.

Research focus

I build dependable Edge AI by making reliability measurable at decision time under hard compute budgets and distribution shift.

Most ML pipelines still lean on two convenient assumptions: near-i.i.d. evaluation and unconstrained inference. In deployed edge autonomy, both assumptions break. My work designs for the real regime directly by co-designing:
(i) dependence-aware reliability signals, (ii) computation-aware inference controllers, and (iii) protocol-driven validation so systems can report calibrated uncertainty and cost, defer when warranted, and avoid silent failure.

Active projects and proof points

NSF CPS AERIAL (Co-PI; Thrust 2 Lead): I lead budget-aware learning-and-control and digital-twin validation protocols that report reliability in operational units (e.g., constraint violations and energy-per-decision) using replayable traces, targeting UAV swarm autonomy in a search-and-rescue setting.
TinyGNN + Low-rank training for deployable graph inference: I develop compact graph inference primitives that make decision-time cost controllable. A low-rank message-passing architecture achieves up to 60× model size reduction with only ~2% best-case performance drop, enabling HW/SW co-design via tunable rank/precision knobs.
Budgeted test-time inference for foundation models (CGES): We co-developed Confidence-Guided Early Stopping, a Bayesian stopping rule that uses confidence as evidence and halts sampling when posterior mass crosses a threshold (or a hard budget binds). Across five reasoning benchmarks, CGES reduces model calls by 69.4% while matching self-consistency accuracy.
Foundations: minimax sample complexity for GNNs under dependence/topology: I develop minimax limits that characterize when graph topology makes learning intrinsically label-inefficient, yielding diagnostics that predict when more labels help versus when gains require changing mixing geometry.

Future research directions

My program advances a single goal: dependable Edge AI under dependence and hard budgets, through three composable thrusts:

Computation-aware inference controllers (adaptive rank/precision/sampling depth with explicit accuracy–efficiency tradeoffs).
Protocol-driven validation + lightweight monitoring (replayable stress tests, operational metrics like Joules/decision, and label-free drift monitors). {index=9}
Dependence-aware learning foundations (evaluation and uncertainty for relational/time-varying streams; topology-aware diagnostics).

Teaching (and why students don’t hate it)

I teach ML, generative models, DSP, image processing, and data science foundations. My goal is to turn technical knowledge into durable engineering ability: students learn to specify requirements, build baselines, debug under shift/partial observability, and deliver reproducible artifacts that justify accuracy–reliability–cost tradeoffs. My UMass evaluations reflect strong execution and clarity (overall instructor effectiveness 4.7/5.0, with 5.0/5.0 on preparation and use of class time in a recent offering).

Students, collaborators, and funding

I am building research around shared interfaces that make systems dependable in practice: reliability contracts, telemetry/monitoring, and trace-based evaluation suites spanning Networking/IoT, embedded/SoC, and autonomy.
I currently mentor PhD and undergraduate researchers and regularly collaborate across ML systems and wireless/autonomy.

If you are interested in dependable ML systems for edge autonomy, resource-aware graph learning, or budgeted inference for foundation models, feel free to reach out.

news

Jul 15, 2025	Our NSF proposal, titled “AERIAL (AI-Embedded Responsive Intelligent Agents with Trajectory-Induced Digital Twin Learning” is awarded by National Science Foundation (NSF).
Feb 16, 2024	Our paper on Tiny Graph Neural Network has been accepted for presentation at the 2024 TinyML Research Symposium!
Jan 17, 2024	Our paper is accepted to IEEE Transactions on Vehicular Technology.
Oct 1, 2023	I served as a reviewer for IEEE Transactions on Wireless Communications.
Jul 1, 2022	I received Travel Grant Award from School of Arts & Sciences, WPI, to present my paper at IEEE AP-S/URSI 2022!
May 22, 2022	Our paper is accepted to IEEE AP-S/URSI 2022.
Nov 1, 2021	I served as a reviewer for IEEE Transactions on Wireless Communications and IEEE Transactions on Communications.

selected publications

Adversarial Attacks on Graph Neural Networks based Spatial Resource Management in P2P Wireless Communications

Ahmad Ghasemi, Ehsan Zeraatkar, Majid Moradikia, and 2 more authors

2023

Abs arXiv Bib HTML

This paper introduces adversarial attacks targeting a Graph Neural Network (GNN) based radio resource management system in point to point (P2P) communications. Our focus lies on perturbing the trained GNN model during the test phase, specifically targeting its vertices and edges. To achieve this, four distinct adversarial attacks are proposed, each accounting for different constraints, and aiming to manipulate the behavior of the system. The proposed adversarial attacks are formulated as optimization problems, aiming to minimize the system’s communication quality. The efficacy of these attacks is investigated against the number of users, signal-to-noise ratio (SNR), and adversary power budget. Furthermore, we address the detection of such attacks from the perspective of the Central Processing Unit (CPU) of the system. To this end, we formulate an optimization problem that involves analyzing the distribution of channel eigenvalues before and after the attacks are applied. This formulation results in a Min-Max optimization problem, allowing us to detect the presence of attacks. Through extensive simulations, we observe that in the absence of adversarial attacks, the eigenvalues conform to Johnson’s SU distribution. However, the attacks significantly alter the characteristics of the eigenvalue distribution, and in the most effective attack, they even change the type of the eigenvalue distribution.
@misc{ghasemi2023adversarial, title = {Adversarial Attacks on Graph Neural Networks based Spatial Resource Management in P2P Wireless Communications}, author = {Ghasemi, Ahmad and Zeraatkar, Ehsan and Moradikia, Majid and Seyed and Zekavat}, year = {2023}, eprint = {2312.08181}, archiveprefix = {arXiv}, primaryclass = {eess.SP}, doi = {10.1109/TVT.2024.3360145}, }
IEEE WiSEE
Adversarial Attacks on Resource Management in P2P Wireless Communications

Ahmad Ghasemi, Ehsan Zeraatkar, Majid Moradikia, and 1 more author

In 2023 IEEE International Conference on Wireless for Space and Extreme Environments (WiSEE), 2023

Abs Bib HTML Code

This paper explores adversarial attacks on a Graph Neural Network (GNN) based radio resource management in point-to-point (P2P) communications. The trained GNN model, which receives information from transceiver pairs, is targeted during the test phase. The paper introduces a novel adversarial attack that modifies the vertices of the GNN model, taking into account various constraints. The attack’s effectiveness is evaluated based on the number of users and signal-to-noise ratio (SNR). The proposed attack formulates optimization problems aimed at minimizing system communication quality, incorporating specific constraints.
@inproceedings{10289604, author = {Ghasemi, Ahmad and Zeraatkar, Ehsan and Moradikia, Majid and Zekavat, Seyed Reza}, booktitle = {2023 IEEE International Conference on Wireless for Space and Extreme Environments (WiSEE)}, title = {Adversarial Attacks on Resource Management in P2P Wireless Communications}, year = {2023}, volume = {}, number = {}, pages = {148-153}, doi = {10.1109/WiSEE58383.2023.10289604}, dimensions = {true}, }
IEEE USNC-URSI
On Eigenvalue Distribution of Imperfect CSI in mmWave Communications

Ahmad Ghasemi, and Seyed Reza Zekavat

In 2022 IEEE USNC-URSI Radio Science Meeting (Joint with AP-S Symposium), 2022

Abs Bib HTML

Imperfect Channel State Information (CSI) reduces communication performance of mmWave communication that uses massive antennas. The knowledge of eigenvalue distribution of the imperfectly estimated CSI between transmitter (TX) and users’ antennas is key to techniques that suppress the impact of imperfect CSI on communication performance. This paper numerically depicts that this distribution follow a power-law distribution. In addition, this paper investigates the impact of channel estimation error, the number of antenna elements at TX and users, and the number of users on the eigen value distribution.
@inproceedings{9887493, author = {Ghasemi, Ahmad and Zekavat, Seyed Reza}, booktitle = {2022 IEEE USNC-URSI Radio Science Meeting (Joint with AP-S Symposium)}, title = {On Eigenvalue Distribution of Imperfect CSI in mmWave Communications}, year = {2022}, volume = {}, number = {}, award = {Travel Award}, pages = {56-57}, doi = {10.23919/USNC-URSI52669.2022.9887493}, }
IEEE TWC
Low-Cost mmWave MIMO Multi-Streaming via Bi-Clustering, Graph Coloring, and Hybrid Beamforming

Ahmad Ghasemi, and Seyed A. Zekavat

IEEE Transactions on Wireless Communications, 2021

Abs Bib HTML

This paper proposes and analyses a set of novel and optimum multi-streaming techniques for mmWave multi-input multi-output systems. It formulates an optimization problem that enables exploiting available uncorrelated paths between the transmitter (Tx) and the receiver (Rx) to enhance the system throughput. In the proposed approach, antenna arrays at Tx/Rx are modeled as a Bipartite graph. Next, two bi-clustering algorithms are applied to the graph to simultaneously cluster antenna elements at Tx and Rx. In addition, the paper shows how the selection of subchannels and their corresponding subantenna arrays can be reduced to a variant of the graph coloring problem, based on which, two algorithms are proposed to find the optimum subchannels and subantenna arrays. Moreover, the paper defines two new beamforming methods and proves that those methods satisfy constant modulus and total power constraints. The first modified beamforming method uses singular vectors (SVs) of subchannels between Tx/Rx and incorporates the Power Iteration algorithm to decrease singular value decomposition complexity. The second newly proposed beamforming method finds precoders/combiners without using SVs which reduces the computational complexity. Performance evaluations in terms data streaming sum-rate demonstrate that the proposed technique increases the throughput using a low processing complexity.
@article{9351753, author = {Ghasemi, Ahmad and Zekavat, Seyed A.}, journal = {IEEE Transactions on Wireless Communications}, title = {Low-Cost mmWave MIMO Multi-Streaming via Bi-Clustering, Graph Coloring, and Hybrid Beamforming}, year = {2021}, volume = {20}, number = {7}, pages = {4113-4127}, doi = {10.1109/TWC.2021.3056077}, }
IEEE WOCC
Joint Hybrid Beamforming and Dynamic Antenna Clustering for Massive MIMO

Ahmad Ghasemi, and Seyed Reza Zekavat

In 2020 29th Wireless and Optical Communications Conference (WOCC), 2020

Abs Bib HTML

This paper offers a new approach for antenna clustering and hybrid beamforming applicable to massive MIMO systems. Simultaneous clustering and hybrid beamforming across Tx and Rx antennas is an NP-hard problem. To address this issue, first, the paper proposes an antenna clustering that is applied to both Tx and Rx. In this regard, antenna arrays at Tx and Rx are modeled as a Bipartite graph and for the first time, one bi-clustering algorithm, Spectral Co-Clustering algorithm, is applied to achieve simultaneous clustering. Next, singular vectors of subchannels, which are the channels between subantenna arrays of Tx and Rx, are comprised to determine optimal precoders/combiners. Performance evaluations in terms of Tx-Rx data streaming sum-rate demonstrate the effectiveness of the proposed algorithm.
@inproceedings{9114913, author = {Ghasemi, Ahmad and Zekavat, Seyed Reza}, booktitle = {2020 29th Wireless and Optical Communications Conference (WOCC)}, title = {Joint Hybrid Beamforming and Dynamic Antenna Clustering for Massive MIMO}, year = {2020}, volume = {}, number = {}, pages = {1-6}, doi = {10.1109/WOCC48579.2020.9114913}, award = {Best Paper Award}, }