Machine learning researcher working on robust and adaptive computer vision and RF/wireless systems.
I work on making deep learning dependable when the world stops matching the training set. My research builds
test-time and online adaptation methods that recover accuracy on the fly under distribution shift — and
studies how those same mechanisms can be attacked, from label-free adversarial perturbations to projected
light on autonomous lane-centering systems. The work spans computer vision and RF machine learning, with a
consistent emphasis on what actually runs on edge hardware.
News
Paper
FIDEI: Hyperparameter-Free Out-of-Distribution Detection accepted at WACV 2027.
Patent
Our U.S. patent application on DARDA — domain-aware real-time dynamic adaptation of
edge-assisted neural networks — was allowed for issuance.
Paper
Optical Adversaries: Assessing Robustness of Test-Time Adaptation Against Light-Based Adversarial
Perturbation accepted at SPIE Assurance and Security for AI-enabled Systems 2026.
Paper
Metacognitive Artificial Intelligence in Vision Foundation Models: Research Challenges to appear
in IEEE Intelligent Systems.
Service
Serving as a reviewer for NeurIPS 2026.
Presentation
On the Adversarial Vulnerability of Label-Free Test-Time Adaptation presented at
ICLR 2025.
Deep networks are trained once and deployed for years into weather, sensor drift, jamming, and adversaries that
were never in the training set. My work asks two questions at the same time: how can a model keep working as the
world moves, and how can that adaptation be broken?
01
Test-time and online domain adaptation
I designed and led development of DARDA, a domain-aware method that adapts a deployed
network in real time and restores accuracy on the fly as conditions change on edge devices. The approach
produced a WACV 2025 publication and a U.S. patent allowed for issuance in 2026. Related work extends
adaptation to tactical wireless systems in a zero-shot setting, and to semantic segmentation through
shift-aware, affinity-guided updates.
WACV 2025MILCOM 2023Under review
02
The security of adaptation itself
Label-free test-time adaptation updates a model using unlabeled test data — which means the test stream is a
control channel. At ICLR 2025 we showed how this exposes a practical attack surface, and in
RF machine learning systems we characterized adversarial dynamic adaptation under realistic constraints. The
broader goal is to build experimental frameworks that stress-test the reliability of adaptation before it
reaches safety-critical deployment.
ICLR 2025MILCOM 2024
03
Physical and light-based attacks on perception
Digital perturbations rarely survive the trip through a camera. I study perturbations delivered by light —
projected patterns and optical interference — and what they do to adaptive perception stacks, including lane
centering systems in autonomous driving.
SPIE 2026Under review
04
Out-of-distribution detection and metacognition
A model that knows when it does not know is worth more than a marginally more accurate one. This line of work
covers hyperparameter-free OOD detection, a comprehensive survey of OOD detection in computer vision, and
research challenges for metacognitive behavior in vision foundation models.
WACV 2027IEEE Intelligent Systems 2026Survey
Collaborations & Sponsors
This research has been carried out across multi-stakeholder programs, where I coordinated timelines and
technical deliverables spanning academic, government, and industry collaborators.
DARPA
National Science Foundation
Army Research Laboratory
Air Force Research Laboratory
Technology Innovation Institute, Abu Dhabi
Publications
Papers, preprints & patents
Eleven peer-reviewed publications and a U.S. patent allowed for issuance. My name is
underlined in each author list.
Under Review
Review
SAGE: Shift Aware Affinity Guided Test Time Adaptation for Semantic Segmentation
Shahriar Rifat, Lorenzo Bellone, Jennifer Simonjan, Francesco Restuccia
Under review, 2026
Review
LaneMirage: Projected-Light Attack on Autonomous Lane Centering Systems
Shahriar Rifat, Nathaniel D. Bastian, Francesco Restuccia
DARDA: Domain-Aware Real-Time Dynamic Adaptation of Edge-Assisted Neural Networks
Shahriar Rifat, Francesco Restuccia, Jonathan D. Ashdown, Kurt Turck
U.S. Patent Application US 20250111667 · filed Sep 4, 2024 · allowed for issuance Jun 15, 2026
Experience
Research, teaching & service
Research
Jun 2022 — Present
Boston, USA
Graduate Research Assistant
Institute for Intelligent Networked Systems, Northeastern University
Analyzed why deep neural networks fail under distribution shift and adversarial conditions — including
label-free and light-based perturbation attacks — across computer vision and RF/wireless systems,
pinpointing vulnerabilities that inform more robust real-world deployment.
Designed and led development of DARDA, a domain-aware method that adapts neural networks in real time and
restores accuracy on the fly as conditions change on edge devices; produced a WACV 2025 publication and a
U.S. patent allowed for issuance in 2026.
Built and ran experimental frameworks to stress-test the security and reliability of adaptation methods,
extending results across vision and RF machine learning and contributing to 11 peer-reviewed publications.
Managed timelines and cross-team technical coordination for multi-stakeholder research spanning DARPA,
NSF, the Army Research Laboratory, AFRL, and a one-year industry partnership with TII (Abu Dhabi).
Synthesized research on AI reliability, robustness, and security into technical papers, survey articles,
and patent filings for both specialist and interdisciplinary audiences.
Bangladesh University of Engineering and Technology
Teaching
Feb 2020 — Present
On leave
Lecturer, Dept. of CSE
University of Information Technology and Sciences, Dhaka
Taught undergraduate courses in Linear Algebra, Machine Learning, and Robotics; designed assignments and
examinations, held office hours, and mentored students on coursework and projects.
Aug — Dec 2019
Dhaka, Bangladesh
Adjunct Lecturer, Dept. of EEE
Sonargaon University, Dhaka
Taught undergraduate courses in Linear Systems and Digital Signal Processing; developed lecture material and
assessments.
Peer Review
Journals
IEEE Transactions on Cognitive Communications and Networking
I lead by shaping the team environment — clear communication, reliability, and follow-through — building durable
trust across multi-institution research collaborations.
Sustained multi-year collaborations with DARPA, NSF, the Army Research Laboratory, AFRL, and TII (Abu Dhabi),
building the cross-institutional trust needed to carry joint technical work through to publication and patent
filing.
Secretary, IEEE EMBS Bangladesh Chapter (Jan 2020 — Feb 2021)
Conference Secretary, IEEE Region 10 Symposium (TENSYMP) 2020
Conference Secretary, IEEE BECITHCON 2019 and IEEE Co-Located Conferences 2021
Key Organizing Member, IEEE WIECON-ECE 2020 and ICECE 2020
Curriculum Vitae
Full CV & technical profile
The complete CV covers publications, patent filings, funded collaborations, peer review, teaching, and service.
