Subangkar portrait Subangkar

Subangkar Karmaker Shanto

CS PhD Candidate, Purdue University, Indiana, United States of America

West Lafayette, IN, USA

Photo of Subangkar Karmaker Shanto

Hello and thank you for stopping by!

Welcome to my corner of the web. I'm currently a Graduate Teaching Assistant in the Department of Computer Science at Purdue University, where I have the privilege of working in the Cyber2s Lab under the guidance of the esteemed Dr. Elisa Bertino. Prior to this, I worked as a Lecturer at the Department of CSE, United International University, Bangladesh and an adjunct faculty at Bangladesh University of Engineering and Technology (BUET).

I graduated from the Department of Computer Science and Engineering at BUET, an experience that laid the foundation for my passion for technology and innovation. My research interests span a variety of exciting fields, including Agent runtime verification and security, RL, Systems and Network Security and Data Mining.

Beyond the world of code and algorithms, I love immersing myself in new experiences whether it's traveling to unexplored destinations, savoring exotic cuisines, meeting fascinating people, or embracing diverse cultures. Feel free to browse around, learn more about my work, and reach out if you'd like to connect or collaborate, I'd love to hear from you!

Agent Runtime Verification LLM/LVLM Security RL & Deep Learning 5G/O-RAN Security Healthcare ML
Last Updated: September 7, 2026

Education

Purdue University

PhD in Computer Science August 2024 – Present

Bangladesh University of Engineering and Technology (BUET)

B.Sc. in Computer Science and Engineering February 2016 – February 2021
  • CGPA: 3.88/4.00 (Ranked 8th in a class of 143 graduating students)
  • Major CGPA: 3.95/4.00

Research Works

Agent Runtime Verification

Active Research
  • Problem: LLM based Agent driven decisions in critical systems lack formal guarantees of correctness and safety
  • Approach: Designing a runtime verification framework to monitor and validate LLM based Agent tool-calling. Surveying formal methods and monitoring approaches for agentic tool call validation in decision-making with respect to set policies
  • Goal: Ensure correctness and safety of LLM based Agent driven decisions in critical infrastructure through lightweight, deployable runtime checks
  • Stack: LangChain, Python, LLM APIs

Trajectory-Level Security for Agentic AI

Published
  • Collaborators: Alireza Lotfi, Dr. Imtiaz Karim, Dr. Elisa Bertino
  • Problem: LLM based agents execute consequential tasks under operational, organizational, and regulatory constraints, but security is still enforced through isolated per action checks that cannot capture system level violations
  • Contribution: Charted the security landscape across the agentic stack: untrusted prompts, memory, retrieved knowledge, and tool interfaces at the single agent level; identity, trust, capability control, and decision transparency under multi agent delegation; and manipulation of the model routing and execution control plane
  • Impact: Argues that security must become a verifiable property of agent architectures, protocols, and runtimes, shifting assurance from single actions to complete execution trajectories, and lays out open problems in behavioral containment, supply chain integrity, provenance, accountability, and observability
ACM AI Leadership Summit 2026, Visionary Track
arXiv

Breaking 5G on the Lower Layer

Published
  • Collaborators: Dr. Imtiaz Karim, Dr. Elisa Bertino
  • Problem: 5G PHY and MAC lack integrity protection during early access. This exposes real but understudied attack surfaces
  • Approach: Built a 5G testbed with rogue gNodeBs and real phones. Evaluated SIB1 spoofing and designed a Timing Advance manipulation attack
  • Impact: Caused denial of service and battery drain on commercial devices. Showed small Timing Advance offsets reliably trigger radio link failure
  • Stack: C/C++, srsRAN, USRP B210
FutureG 2026 @ NDSS
arXiv

Securing Synchronization in O-RAN Open Fronthaul

Under Review
  • Collaborators: Yiwei Zhang, Dr. Imtiaz Karim, Dr. Elisa Bertino
  • Problem: The O-RAN Open Fronthaul interface disaggregates DU and RU over untrusted networks, exposing time synchronization protocols to adversarial manipulation and timing attacks
  • Contribution: Implementation within the srsRAN codebase to enable time synchronization between DU and RU over the Open Fronthaul interface
  • Impact: Identified vulnerabilities in open fronthaul timing and proposed countermeasures to harden synchronization against spoofing and delay attacks in O-RAN deployments
  • Stack: C/C++, srsRAN, PTP/IEEE 1588, Linux PTP

BayesBeat: Reliable Atrial Fibrillation Detection from Noisy Photoplethysmography Data

Published
  • Collaborators: Sarkar Snigdha Sarathi Das, Masum Rahman, Md. Saiful Islam, Dr. Atif Hasan Rahman, Mohammad Mehedy Masud, Dr. Mohammed Eunus Ali
  • Problem: PPG signals are noisy from motion; predictions need calibrated uncertainty for reliable AF detection.
  • Approach: Bayesian deep learning model in PyTorch that outputs uncertainty alongside predictions.
  • Impact: +7–25% over prior SOTA on the largest public PPG dataset and +10–14% on MIMIC-III; first BDL application in this domain.
Paper Code arXiv

Contrastive Learning Based Approach for Patient Similarity

Active Research
  • Supervisor: Dr. Mohammed Eunus Ali, Dr. Atif Hasan Rahman
  • Problem: Measure patient similarity directly from physiological (PPG) signals with limited labeled data.
  • Approach: Designed a new contrastive loss; conducted an AF case study under data scarcity.
  • Impact: First application of contrastive similarity learning in this domain; preprint available.
arXiv

Publications

Securing Agentic AI: From Per-Action Checks to Trajectory Assurance

Vision Paper
Alireza Lotfi*, Subangkar Karmaker Shanto*, Imtiaz Karim, Elisa Bertino.

ACM AI Leadership Summit 2026, Visionary Track
* Joint first author with equal contribution

Autonomous agents are increasingly used to execute consequential tasks in environments governed by operational constraints, organizational policies, regulatory requirements, and technical standards. Their safety is therefore determined not by the correctness of individual actions, but by whether their overall behavior remains consistent with the rules and invariants of the systems in which they operate. As large language model (LLM)-based agents become more autonomous and increasingly delegate tasks across organizational boundaries, securing them evolves from a single challenge into a broad and interconnected landscape spanning the entire agentic stack. At the single-agent level, untrusted inputs through prompts, memory, retrieved knowledge, and tool interfaces create attack surfaces. In multi-agent settings, delegation and communication introduce challenges related to identity, trust, capability control, and decision transparency, while the underlying model routing and execution control plane remains vulnerable to manipulation and to unverified model provenance. Perhaps the most fundamental challenge is behavioral containment: sequences of individually permissible actions may collectively violate system-level constraints and safety invariants. At the broader level, supply-chain integrity, provenance, accountability, and end-to-end observability remain largely open problems. A common principle unifies these directions: security must become a verifiable property of the architectures, protocols, and runtimes that govern agent behavior, rather than an optional layer of guidance. Charting these challenges provides a roadmap toward trustworthy autonomous agent deployment.
arXiv

Breaking 5G on the Lower Layer

Workshop
Subangkar Karmaker Shanto, Imtiaz Karim, Elisa Bertino.

[FutureG 2026] Workshop on Security and Privacy of Next-Generation Networks, co-located with NDSS 2026

As 3GPP systems have strengthened security at the upper layers of the cellular stack, plaintext PHY and MAC layers have remained relatively understudied, though interest in them is growing. In this work, we explore lower-layer exploitation in modern 5G, where recent releases have increased the number of lower-layer control messages and procedures, creating new opportunities for practical attacks. We present two practical attacks and evaluate them in a controlled lab testbed. First, we reproduce a SIB1 spoofing attack to study manipulations of unprotected broadcast fields. By repeatedly changing a key parameter, the UE is forced to refresh and reacquire system information, keeping the radio interface active longer than necessary and increasing battery consumption. Second, we demonstrate a new Timing Advance (TA) manipulation attack during the random access procedure. By injecting an attacker-chosen TA offset in the random access response, the victim applies incorrect uplink timing, which leads to uplink desynchronization, radio link failures, and repeated reconnection loops that effectively cause denial of service. Our experiments use commercial smartphones and open-source 5G network software. Experimental results in our testbed demonstrate that TA offsets exceeding a small tolerance reliably trigger radio link failures in our testbed and can keep devices stuck in repeated re-establishment attempts as long as the rogue base station remains present. Overall, our findings highlight that compact lower-layer control messages can have a significant impact on availability and power, and they motivate placing defenses for initial access and broadcast procedures.
arXiv

BayesBeat: Reliable Atrial Fibrillation Detection from Noisy Photoplethysmography Data

Journal
Sarkar Snigdha Sarathi Das, Subangkar Karmaker Shanto, Masum Rahman, Md. Saiful Islam, Atif Rahman, Mohammad Mehedy Masud, Mohammed Eunus Ali.

[UbiComp 2022] Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 6, 1, Article 8 (March 2022)
Joint First Author with equal contribution

Smartwatches or fitness trackers have garnered a lot of popularity as potential health tracking devices due to their affordable and longitudinal monitoring capabilities. To further widen their health tracking capabilities, in recent years researchers have started to look into the possibility of Atrial Fibrillation (AF) detection in real-time leveraging photoplethysmography (PPG) data, an inexpensive sensor widely available in almost all smartwatches. A significant challenge in AF detection from PPG signals comes from the inherent noise in the smartwatch PPG signals. In this paper, we propose a novel deep learning based approach, BayesBeat that leverages the power of Bayesian deep learning to accurately infer AF risks from noisy PPG signals, and at the same time provides an uncertainty estimate of the prediction. Extensive experiments on two publicly available dataset reveal that our proposed method BayesBeat outperforms the existing state-of-the-art methods. Moreover, BayesBeat is substantially more efficient having 40-200X fewer parameters than state-of-the-art baseline approaches making it suitable for deployment in resource constrained wearable devices.
Paper Code arXiv

Work Experience

Graduate Teaching Assistant

Department of CS, Purdue University Present
  • Courses Conducted: Programming-I (Java), Project Management, Database Systems

Graduate Research Assistant

Department of CS, Purdue University Fall'24, Summer'25, Summer'26
  • Working in Cyber2s lab under supervision of Dr. Elisa Bertino

Lecturer

Department of CSE, United International University (UIU) February 2021 – August 2024
  • Working as a Lecturer at Department of CSE, United International University, Bangladesh.
  • Courses Conducted: Structured Programming Language, Object Oriented Programming, Artificial Intelligence, Bioinformatics, Algorithms, Computer Networks lab, Human Computer Interaction and Society, Technology & Engineering Ethics

Lecturer (Part-time)

Department of CSE, BUET January 2022 – April 2022
  • Courses Conducted: Structured Programming Language Lab

Research Assistant (Part-time)

Department of CSE, BUET February 2021 – January 2022
  • Worked under supervision of Dr. Atif Hasan Rahman and Dr. Mohammed Eunus Ali in CSE, BUET. Multiple research projects are funded by the government of Bangladesh.

Skills

Technical Skills
  • Programming Languages: Python, C, C++, Java, Assembly Language (Intel x86 Architecture, MIPS Architecture)
  • Scripting Languages: Bash, HTML, CSS, LATEX, SQL
  • Algorithms & Data Structures: 180+ LeetCode problems solved in Python, C++, Java and SQL, with all solutions maintained in a public repository
  • ML & NLP: LLMs, RL, RAG
  • Agentic Systems: Claude Code, OpenAI Codex, Model Context Protocol (MCP), LangChain, tool and function calling, RAG, Benchmarking
  • LLM Serving & Inference: vLLM, Ollama, LiteLLM, Hugging Face Transformers, LLM APIs
  • ML Frameworks: PyTorch, Keras, Gymnasium, Scikit-learn, Pandas
  • Web Frameworks: Django, Django REST Framework
  • Other Frameworks: Corda (Blockchain), JavaFX (Java GUI)
  • SQL Database: Oracle SQL, MySQL, PostgreSQL, SQLite
  • Design Tools: Proteus circuit simulator, Logisim circuit simulator and CISCO packet tracer
  • Hardware Tools: Atmega32 Microcontroller
  • Software & Project Management: Git, GitHub, Docker, Docker Hub
  • Documentation Tools: OpenAPI Specification, Swagger, Redoc
  • IDEs & Editors: PyCharm, CLion, IntelliJ Idea, Codeblocks, VSCode, MS Visual Studio, Tizen Studio, Atmel Studio
  • Software Tools: MS Word, PowerPoint, Excel
LeetCode Solutions
Communication Skills
  • English
  • Bengali (Bangla)
  • Hindi (Limited Proficiency)

Software & Hardware Projects

LVLM / LLM Jailbreaking via Reinforcement Learning

  • Focus: Adaptive jailbreak attacks against LLMs and VLMs that bypass safety alignment.
  • Methods: Reinforcement learning based prompt optimization, PPO, Actor Critic, LLM as judge
  • Impact: Improved attack success rate by 12 to 13 percent with strong cross model transferability
  • Stack: Python, PyTorch, Gymnasium, Ollama, HuggingFace
Code

Security Analysis of 5G Control-Plane Protocols

  • Focus: PHY/MAC-layer analysis to uncover control-plane flaws.
  • Testbed: Mini base-station by modifying open-source radio stacks; async message injection to phones.
  • Impact: Triggered radio link failure, reconnection loops, and battery drain on commercial 5G devices.
  • Stack: C/C++, Java, Open5GS, srsRAN, OpenAirInterface.

Implementation of Precision Time Protocol in srsRAN

  • Focus: Enabling precise time synchronization for 5G O-RAN Fronthaul interface.
  • Goal: Implement PTP (IEEE 1588) between DU and RU in srsRAN for synchronization secuity analysis.
  • Impact: Supports O-RAN compliant deployments and enables accurate timing for 5G research testbeds.
  • Stack: C/C++, srsRAN, PTP/IEEE 1588, Linux PTP.

AI-Generated Text Detection via Adversarial Training

  • Method: distilBERT detector vs. T5-small paraphraser (PPO + back-translation/lexical rewrites).
  • Engineering: ~4× training speed-up via cached corpus generation & PyTorch DataParallel.
Code

SDN Per-Flow Delay/Jitter Prediction with Graph Neural Networks

  • Focus: Graph-based models to predict per-flow delay/jitter.
  • Setup: ONOS + Mininet on AWS; 540 labeled simulations; D-ITG traffic & 27 routing matrices.
Code

FoodSquare — Multi-Tenant Restaurant Marketplace

  • Goal: End-to-end Django platform for restaurants & customers (self-service menus, real-time search/checkout).
  • Deploy: Containerized; one-command deploy to Docker Hub.
Code Docker

Image Captioning in PyTorch

  • Model: ResNet-101 encoder + LSTM with Attention on Flickr8k.
Code

Tizen Native App & Background Service (Sensor Data Collection)

  • Goal: Collect raw sensor data on-device for downstream analysis.
BG Service Native UI App

Real-Time Audio → Frequency Spectrum (Atmega32)

  • Summary: Time-to-frequency conversion and visualization from microphone input on AVR.
Code

Interests

Research Interests
  • LLMs, Agent
  • Systems and Network Security
  • Machine Learning & Data Mining
  • Cellular Security
Hobbies and Other Interests
  • Traveling
  • Watching movies
  • Listening to music

Awards & Services

Achievements
  • Winner: Blockchain Olympiad Bangladesh — February 2021
  • Winner: National Hackathon on Frontier Technologies — February 2020
  • Merit Award Winner: International Blockchain Olympiad — June 2020
  • B.Sc.: Secured place in Dean's List in 3 levels
Other Services
  • Supervisor of Gold Prize Winner team of International Blockchain Olympiad 2023 Final: Hosted in Amsterdam, The Netherlands, EU from 15th to 17th of November 2023. Supervised Team Apocalypse from UIU, Bangladesh. Certificates Link
  • Problem setter of UIU Intra University Deep Learning Sprint Fall 2022: Prepared Dataset to Distinguish between Relevant/Irrelevant Image Captions using Deep Learning. Kaggle Contest Link

Connect with Me

Lawson Computer Science Building, 305 N University St, West Lafayette, IN 47907, USA

View on Google Maps